Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
ANALYSIS
This Page Intentionally Left Blank
INTRODUCTORY
ANALYSIS
A Deeper View of Calculus
Richard J. Bagby
Department of Mathematical Sciences
New Mexico State University
Las Cruces, New Mexico
San Diego San Francisco New York Boston London Toronto Sydney Tokyo
Sponsoring Editor Barbara Holland
Production Editor Julie Bolduc
Editorial Coordinator Karen Frost
Marketing Manager Marianne Rutter
Cover Design Richard Hannus, Hannus Design Associates
Copyeditor Amy Mayﬁeld
Composition TeXnology, Inc./MacroT
E
X
Printer MapleVail Book Manufacturing Group
This book is printed on acidfree paper. ∞
Copyright c 2001 by Academic Press
All rights reserved. No part of this publication may be reproduced or
transmitted in any form or by any means, electronic or mechanical, includ
ing photocopy, recording, or any information storage and retrieval system,
without permission in writing from the publisher.
Requests for permission to make copies of any part of the work should
be mailed to: Permissions Department, Harcourt, Inc., 6277 Sea Harbor
Drive, Orlando, Florida, 328876777.
ACADEMIC PRESS
A Harcourt Science and Technology Company
525 B Street, Suite 1900, San Diego, CA 921014495, USA
http://www.academicpress.com
Academic Press
Harcourt Place, 32 Jamestown Road, London NW1 7BY, UK
Harcourt/Academic Press
200 Wheeler Road, Burlington, MA 01803
http://www.harcourtap.com
Library of Congress Catalog Card Number: 00103265
International Standard Book Number: 0120725509
Printed in the United States of America
00 01 02 03 04 MB 9 8 7 6 5 4 3 2 1
CONTENTS
ACKNOWLEDGMENTS ix
PREFACE xi
I
THE REAL NUMBER SYSTEM
1. Familiar Number Systems 1
2. Intervals 6
3. Suprema and Inﬁma 11
4. Exact Arithmetic in R 17
5. Topics for Further Study 22
II
CONTINUOUS FUNCTIONS
1. Functions in Mathematics 23
2. Continuity of Numerical Functions 28
v
vi CONTENTS
3. The Intermediate Value Theorem 33
4. More Ways to Form Continuous Functions 36
5. Extreme Values 40
III
LIMITS
1. Sequences and Limits 46
2. Limits and Removing Discontinuities 49
3. Limits Involving ∞ 53
IV
THE DERIVATIVE
1. Differentiability 57
2. Combining Differentiable Functions 62
3. Mean Values 66
4. Second Derivatives and Approximations 72
5. Higher Derivatives 75
6. Inverse Functions 79
7. Implicit Functions and Implicit Differentiation 84
V
THE RIEMANN INTEGRAL
1. Areas and Riemann Sums 93
2. Simplifying the Conditions for Integrability 98
3. Recognizing Integrability 102
4. Functions Deﬁned by Integrals 107
5. The Fundamental Theorem of Calculus 112
6. Topics for Further Study 115
VI
EXPONENTIAL AND LOGARITHMIC FUNCTIONS
1. Exponents and Logarithms 116
2. Algebraic Laws as Deﬁnitions 119
CONTENTS vii
3. The Natural Logarithm 124
4. The Natural Exponential Function 127
5. An Important Limit 129
VII
CURVES AND ARC LENGTH
1. The Concept of Arc Length 132
2. Arc Length and Integration 139
3. Arc Length as a Parameter 143
4. The Arctangent and Arcsine Functions 147
5. The Fundamental Trigonometric Limit 150
VIII
SEQUENCES AND SERIES OF FUNCTIONS
1. Functions Deﬁned by Limits 153
2. Continuity and Uniform Convergence 160
3. Integrals and Derivatives 164
4. Taylor’s Theorem 168
5. Power Series 172
6. Topics for Further Study 177
IX
ADDITIONAL COMPUTATIONAL METHODS
1. L’Hˆ opital’s Rule 179
2. Newton’s Method 184
3. Simpson’s Rule 187
4. The Substitution Rule for Integrals 191
REFERENCES 197
INDEX 198
This Page Intentionally Left Blank
ACKNOWLEDGMENTS
I
would like to thank many persons for the support and assistance that
I have received while writing this book. Without the support of my
department I might never have begun, and the feedback I have received
from my students and from reviewers has been invaluable. I would espe
cially like to thank Professors William Beckner of University of Texas at
Austin, Jung H. Tsai of SUNY at Geneseo and Charles Waters of Mankato
State University for their useful comments. Most of all I would like to
thank my wife, Susan; she has provided both encouragement and impor
tant technical assistance.
ix
This Page Intentionally Left Blank
PREFACE
I
ntroductory real analysis can be an exciting course; it is the gateway to
an impressive panorama of higher mathematics. But for all too many
students, the excitement takes the form of anxiety or even terror; they
are overwhelmed. For many, their study of mathematics ends one course
sooner than they expected, and for many others, the doorways that should
have been opened now seem rigidly barred. It shouldn’t have to be that
way, and this book is offered as a remedy.
GOALS FOR INTRODUCTORY ANALYSIS
The goals of ﬁrst courses in real analysis are often too ambitious. Stu
dents are expected to solidify their understanding of calculus, adopt an
abstract point of view that generalizes most of the concepts, recognize how
explicit examples ﬁt into the general theory and determine whether they
satisfy appropriate hypotheses, and not only learn deﬁnitions, theorems,
and proofs but also learn how to construct valid proofs and relevant exam
ples to demonstrate the need for the hypotheses. Abstract properties such
as countability, compactness and connectedness must be mastered. The
xi
xii PREFACE
students who are up to such a challenge emerge ready to take on the world
of mathematics.
A large number of students in these courses have much more modest
immediate needs. Many are only interested in learning enough mathemat
ics to be a good highschool teacher instead of to prepare for highlevel
mathematics. Others seek an increased level of mathematical maturity,
but something less than a quantum leap is desired. What they need is a
new understanding of calculus as a mathematical theory — how to study
it in terms of assumptions and consequences, and then check whether the
needed assumptions are actually satisﬁed in speciﬁc cases. Without such an
understanding, calculus and real analysis seem almost unrelated in spite of
the vocabulary they share, and this is why so many good calculus students
are overwhelmed by the demands of higher mathematics. Calculus stu
dents come to expect regularity but analysis students must learn to expect
irregularity; real analysis sometimes shows that incomprehensible levels
of pathology are not only possible but theoretically ubiquitous. In calcu
lus courses, students spend most of their energy using ﬁnite procedures
to ﬁnd solutions, while analysis addresses questions of existence when
there may not even be a ﬁnite algorithm for recognizing a solution, let
alone for producing one. The obstacle to studying mathematics at the next
level isn’t just the inherent difﬁculty of learning deﬁnitions, theorems, and
proofs; it is often the lack of an adequate model for interpreting the abstract
concepts involved. This is why most students need a different understand
ing of calculus before taking on the abstract ideas of real analysis. For
some students, such as prospective highschool teachers, the next step in
mathematical maturity may not even be necessary.
The book is written with the future teacher of calculus in mind, but it is
also designed to serve as a bridge between a traditional calculus sequence
and later courses in real or numerical analysis. It provides a viewof calculus
that is now missing from calculus books, and isn’t likely to appear any time
soon. It deals with derivations and justiﬁcations instead of calculations and
illustrations, with examples showing the need for hypotheses as well as
cases in which they are satisﬁed. Deﬁnitions of basic concepts are empha
sized heavily, so that the classical theorems of calculus emerge as logical
consequences of the deﬁnitions, and not just as reasonable assertions based
on observations. The goal is to make this knowledge accessible without
diluting it. The approach is to provide clear and complete explanations of
the fundamental concepts, avoiding topics that don’t contribute to reaching
our objectives.
PREFACE xiii
APPROACH
To keep the treatments brief yet comprehensible, familiar arguments have
been reexamined, and a surprisingly large number of the traditional con
cepts of analysis have proved to be less than essential. For example, open
and closed intervals are needed but open and closed sets are not, sequences
are needed but subsequences are not, and limits are needed but methods
for ﬁnding limits are not. Another key to simplifying the development is
to start from an appropriate level. Not surprisingly, completeness of the
real numbers is introduced as an axiom instead of a theorem, but the axiom
takes the form of the nested interval principle instead of the existence of
suprema or limits. This approach brings the power of sequences and their
limits into play without the need for a ﬁne understanding of the difference
between convergence and divergence. Suprema and inﬁma become more
understandable, because the proof of their existence explains what their
deﬁnition really means. By emphasizing the deﬁnition of continuity in
stead of limits of sequences, we obtain remarkably simple derivations of
the fundamental properties of functions that are continuous on a closed
interval:
existence of intermediate values
existence of extreme values
uniform continuity.
Moreover, these fundamental results come early enough that there is plenty
of time to develop their consequences, such as the mean value theorem,
the inverse function theorem, and the Riemann integrability of continu
ous functions, and then make use of these ideas to study the elementary
transcendental functions. At this stage we can begin mainstream real anal
ysis topics: continuity, derivatives, and integrals of functions deﬁned by
sequences and series.
The coverage of the topics studied is designed to explain the concepts,
not just to prove the theorems efﬁciently. As deﬁnitions are given they
are explained, and when they seem unduly complicated the need for the
complexity is explained. Instead of the deﬁnition  theorem  proof format
often used in sophisticated mathematical expositions, we try to see how
the deﬁnitions evolve to make further developments possible. The rigor
is present, but the formality is avoided as much as possible. In general,
proofs are given in their entirety rather in outline form; the reader isn’t left
with a sequence of exercises to complete them.
xiv PREFACE
Exercises at the end of each section are designed to provide greater
familiarity with the topics treated. Some clarify the arguments used in
the text by having the reader develop parallel ones. Others ask the reader
to determine how simple examples ﬁt into the general theory, or give
examples that highlight the relevance of various conditions. Still others
address peripheral topics that the reader might ﬁnd interesting, but that
were not necessary for the development of the basic theory. Generally the
exercises are not repetitive; the intent is not to provide practice for working
exercises of any particular type, and so there are few worked examples to
follow. Computational skill is usually important in calculus courses but
that is not the issue here; the skills to be learned are more in the nature of
making appropriate assumptions and working out their consequences, and
determining whether various conditions are satisﬁed. Such skills are much
harder to develop, but well worth the effort. They make it possible to do
mathematics.
ORGANIZATION AND COVERAGE
The ﬁrst seven chapters treat the fundamental concepts of calculus in a rig
orous manner; they form a solid core for a onesemester course. The ﬁrst
chapter introduces the concepts we need for working in the real number
system, and the second develops the remarkable properties of continuous
functions that make a rigorous development of calculus possible. Chapter
3 is a deliberately brief introduction to limits, so that the fundamentals of
differentiation and integration can be reached as quickly as possible. It
shows little more than how continuity allows us to work with quantities
given as limits. The fourth chapter studies differentiability; it includes a
development of the implicit function theorem, a result that is not often pre
sented at this level. Chapter 5 develops the theory of the Riemann integral,
establishing the equivalence of Riemann’s deﬁnition with more convenient
ones and treating the fundamental theorem even when the integrand fails to
be a derivative. The sixth chapter studies logarithms and exponents froman
axiomatic point of view that leads naturally to formulas for them, and the
seventh studies arc length geometrically before examining the connections
between arc length and calculus.
Building on this foundation, Chapter 8 gets into mainstream real anal
ysis, with a deeper treatment of limits so that we can work with sequences
and series of functions and investigate questions of continuity, differentia
bility, and integrability. It includes the construction of a function that is
continuous everywhere but nowhere differentiable or monotonic, showing
that calculus deals with functions much more complicated than we can
PREFACE xv
visualize, and the theory of power series is developed far enough to prove
that each convergent power series is the Taylor series for its sum. The ﬁnal
chapter gives a careful analysis of some additional topics that are commonly
learned in calculus but rarely explained fully. They include L’Hˆ opital’s rule
and an analysis of the error in Simpson’s rule and Newton’s method; these
could logically have been studied earlier but were postponed because they
were not needed for further developments. They could serve as indepen
dent study projects at various times in a course, rather than studied at the
end.
A few historical notes are included, simply because they are interest
ing. While a historical understanding of calculus is also desirable, some
traditional calculus texts, such as the one by Simmons [3], already meet
this need.
GETTING THE MOST FROM THIS BOOK
Books should be read and mathematics should be done; students should ex
pect to do mathematics while reading this book. One of my primary goals
was to make it read easily, but reading it will still take work; a smooth
phrase may be describing a difﬁcult concept. Take special care in learning
deﬁnitions; later developments will almost always require a precise under
standing of just exactly what they say. Be especially wary of unfamiliar
deﬁnitions of familiar concepts; that signals the need to adopt an unfamil
iar point of view, and the key to understanding much of mathematics is to
examine it from the right perspective. The deﬁnitions are sometimes more
complex than they appear to be, and understanding the stated conditions
may involve working through several logical relationships. Each reader
should try to think of examples of things that satisfy the relevant conditions
and also try to ﬁnd examples of things that don’t; understanding how a
condition can fail is a key part of understanding what it really means.
Take the same sort of care in reading the statement of a theorem; the
hypotheses and the conclusion need to be identiﬁed and then understood.
Instead of reading a proof passively, the reader should work through the
steps described and keep track of what still needs to be done, question why
the approach was taken, check the logic, and look for potential pitfalls. A
writer of mathematics usually expects this level of involvement, and that’s
why the word “we” appears so often in work by a single author. With
an involved reader, the mathematics author can reveal the structure of an
argument in a way that is much more enlightening than an overly detailed
presentation would be.
xvi PREFACE
Pay close attention to the role of stated assumptions. Are they made
simply for the purposes of investigation, to make exploration easier, or are
they part of a rigorous argument? Are the assumptions known to be true
whenever the stated hypotheses are satisﬁed, or do they simply correspond
to special cases being considered separately? Or is an assumption made
solely for the sake of argument, in order to show that it can’t be true?
Mastering the material in this book will involve doing mathematics
actively, and the same is probably true of any activity that leads to greater
mathematical knowledge. It is work, but it is rewarding work, and it can
be enjoyable.
I
THE REAL NUMBER SYSTEM
F
rom counting to calculus, the methods we use in mathematics are
intimately related to the properties of the underlying number system.
So we begin our study of calculus with an examination of real
numbers and how we work with them.
1 FAMILIAR NUMBER SYSTEMS
The ﬁrst numbers we learn about are the natural numbers N, which are
just the entire collection of positive integers 1, 2, 3, . . . that we use for
counting. But the natural number system is more than just a collection
of numbers. It has additional structure, and the elements of N can be
identiﬁed by their role in this structure as well as by the numerals we
ordinarily use. For example, each natural number n ∈ N has a unique
successor n
∈ N; we’re used to calling the successor n + 1. No natural
number is the successor of two different natural numbers. The number we
call 1 is the only element of Nthat is not a successor of any other element
1
2 CHAPTER I THE REAL NUMBER SYSTEM
of N. All the other elements of Ncan be produced by forming successors:
2 is the successor of 1, 3 is the successor of 2, 4 is the successor of 3, and
so on.
There’s a wealth of information in the preceding paragraph. It includes
the basis for an important logical principle called mathematical induction.
In its simplest form, it says that if a set of natural numbers contains 1 and
also contains the successor of each of its elements, then the set is all of N.
That allows us to deﬁne operations on None element at a time and provides
us with a powerful method for establishing their properties. For example,
consider the addition of natural numbers. What is meant by m+n? Given
m ∈ N, we can deﬁne the sum m + 1 to be the successor of m, m + 2
to be the successor of m + 1, and so on. Once we’ve deﬁned m + n, we
can deﬁne m plus the successor of n to be the successor of m + n. So
the set of all n for which this process deﬁnes m + n is a set that contains
1 and the successor of each of its elements. According to the principle of
mathematical induction, this process deﬁnes m+n for all natural numbers
n.
If we’re ambitious, we can use this formal deﬁnition of the addition of
natural numbers to develop rigorous proofs of the familiar laws for addition.
This was all worked out by an Italian mathematician, Giuseppe Peano, in
the late nineteenth century. We won’t pursue his development further,
since our understanding of the laws of arithmetic is already adequate for
our purposes. But we will return to the principle of mathematical induction
repeatedly. It appears in two forms: in inductive deﬁnitions, as above, and
in inductive proofs. Sometimes it appears in the form of another principle
we use for working with sets of natural numbers: each nonempty subset of
N contains a smallest element. Mathematicians refer to this principle by
saying that Nis a wellordered set.
To improve our understanding of mathematical induction, let’s use it
to prove that N is wellordered. We need to show that every subset of N
is either empty or has a smallest element. We’ll do so by assuming only
that E ⊂ Nand that E has no smallest element, then proving that E must
be the empty set. Our strategy is prove that n ∈ Nimplies that n / ∈ E, so
that E can’t have any elements. We indicate this by writing E = ∅, the
standard symbol for the empty set. It’s easy to see why 1 / ∈ E; whenever
1 is in a set of natural numbers, it is necessarily the smallest number in the
set. After we’ve learned that 1 / ∈ E, we can deduce that 2 / ∈ E because 2
is the smallest number in any set of natural numbers that contains 2 but not
1. This line of reasoning can be continued, and that’s how induction comes
in. Call I
n
the set {1, 2, . . . , n} with n ∈ Nso that I
n
∩E (the intersection
1. FAMILIAR NUMBER SYSTEMS 3
of I
n
and E) represents all the numbers in E that are included in the ﬁrst
n natural numbers. Whenever n has the property that I
n
∩ E = ∅, we can
use the assumption that E has no smallest element to deduce that n + 1
can’t be in E, and therefore I
n+1
∩ E = ∅ as well. Since we do know
that I
1
∩ E = ∅, the inductive principle guarantees that I
n
∩ E = ∅ for all
n ∈ N.
The next development is to use addition to deﬁne the inverse operation,
subtraction, with m−n initially deﬁned only for mlarger than n. To make
subtraction of arbitrary natural numbers possible, we enlarge the system
N to form Z, the system of all integers (the symbol Z comes from the
German word Zahl for numeral). The system Z includes N, 0, and each
negative integer −n with n ∈ N. As in N, each integer j ∈ Zhas a unique
successor in Z, but the difference is that each has a unique predecessor as
well. If we try to run through the elements of Z one at a time, we see that
we can’t list them in increasing order; there’s no place we can start without
leaving out lower integers. But if we don’t try to put them in numerical
order, there are lots of ways to indicate them in a list, such as
Z = {0, 1, −1, 2, −2, . . . , n, −n, . . .}.
As we do mathematics, we often consider the elements of a set one at a
time, with both the starting point and the progression clearly deﬁned. When
we do so, we are working with a mathematical object called a sequence,
the mathematical termfor a numbered list. On an abstract level, a sequence
in a set S is a special sort of indexed collection {x
n
: n ∈ I} with x
n
∈ S
for each n ∈ I. The indexed collection is a sequence if the index set I is a
subset of Z with two special properties: it has a least element, and for each
element n of I either n + 1 ∈ I or n is the greatest element of I. That lets
us use mathematical induction to deﬁne sequences or to prove things about
speciﬁc ones, as in our proof that N is a wellordered set. In the abstract,
we prefer index sets with 1 the least element, so that x
1
is the ﬁrst element
of the sequence and x
n
is the nth element. Logically, we could just as
well make this preference a requirement. But in practice it may be more
natural to do something else, and that’s the reason we deﬁned sequences
the way we did. For example, we might deﬁne a sequence by calling a
n
the coefﬁcient of x
n
in the expansion of
_
x
−1
+ 1 +x
_
7
, making a
−7
the
ﬁrst term and a
7
the last.
Since the index set of a sequence is completely speciﬁed by giving its
least element and either giving its largest element or saying that it has none,
sequences are often indicated by giving this information about the index
set in place of an explicit deﬁnition of it. For example, {x
n
}
7
n=0
indicates
4 CHAPTER I THE REAL NUMBER SYSTEM
a sequence with index set {0, 1, 2, 3, 4, 5, 6, 7}, and {x
n
}
∞
n=1
indicates a
sequence with index set all of N.
The rational number system Qcontains all fractions m/n with n = 0
and m, n ∈ Z. The symbol Q is used because rational numbers are
quotients of integers. Different pairs of integers can be used to indicate the
same rational number. The rule is that
m
n
=
m
∗
n
∗
if mn
∗
= m
∗
n.
It’s at this stage that the difference between mathematical operations and
other things we might do with mathematical symbols becomes important.
When we perform a mathematical operation on a rational number x, the
result isn’t supposed to depend on the particular pair of integers we use
to represent x, even though we may use them to express the result. For
example, we may add 2 to the rational number
m
n
and express the result
as
2n+m
n
; this is a legitimate mathematical operation. On the other hand,
converting
m
n
to
2 +m
n
doesn’t correspond to any process we can call a
mathematical operation on the rational number
m
n
, it’s just something else
we can do with m and n.
The rational numbers can also be identiﬁed with their decimal expan
sions. When the quotient of two integers is computed using the standard
long division algorithm, the resulting decimal either terminates or repeats.
Conversely, any terminating decimal or repeating decimal represents a ra
tional number; that is, it can be written as the quotient of two integers.
Obviously, a terminating decimal can be written as an integer divided by a
power of 10. There’s also a way to write a repeating decimal as a termi
nating decimal divided by a natural number of the form 10
n
− 1. There’s
a clever trick involved: if the repeating part of x has n digits, then 10
n
x
has the same repeating part, and so the repeating parts cancel out when we
compute (10
n
−1)x as 10
n
x −x.
When considered in numerical order, elements of Q have neither im
mediate successors nor predecessors. No matter which two distinct rational
numbers we specify, there are always additional rational numbers between
them. So we never think of one rational number as being next to another.
If we disregard the numerical order of the rationals, it is still possible to
come up with a sequence that includes every element in Q, but we have no
reason to pursue that here.
It’s common to think of the rational numbers as arranged along the
xaxis in the coordinate plane. The rational number
m
n
corresponds to the
intersection of the xaxis with the line through the lattice points (m, n −1)
and (0, −1). Although it looks like one can ﬁll the entire xaxis with such
EXERCISES 5
points, it’s been known since the time of Pythagoras that some points on
the line don’t correspond to any rational number. For example, if we form
a circle by putting its center at (0, 0) and choosing the radius to make the
circle pass through (1, 1), then the circle crosses the xaxis twice but not
at any point in Q. Extending Qto a larger set of numbers that corresponds
exactly to all the points on the line produces the real number system R.
The fact that there are no missing points is the geometric version of the
completeness property of the real numbers, a property we’ll see a great
deal more of.
Calculus deals with variables that take their values in R, so a reason
ably good understanding of Ris needed before one can even comprehend
the possibilities. The real number system is far more complicated than
one might expect. Our computational algorithms can’t really deal with
complete decimal representations of real numbers. For example, decimal
addition is supposed to begin at the rightmost digit. Mathematicians have
a simple way to get around this difﬁculty; we just ignore it. We sim
ply indicate arithmetic operations in R with algebraic notation, treating
the symbols that represent real numbers in much the same way we treat
unknowns.
EXERCISES
1. Use the principle of mathematical induction to deﬁne 2
n
for all n ∈
N. (A deﬁnition using this principle is usually said to be given
inductively or recursively.)
2. The sequence {s
n
}
∞
n=1
whose nth term is the sum of the squares of
the ﬁrst n natural numbers can be deﬁned using Σnotation as
s
n
=
n
k=1
k
2
.
It can also be deﬁned recursively by specifying
s
1
= 1 and s
n+1
= s
n
+ (n + 1)
2
for all n ∈ N.
Use the inductive principle to prove that s
n
=
1
6
n(n + 1) (2n + 1)
for all n ∈ N.
3. Why is it impossible to ﬁnd a sequence that includes every element
of Z with all the negative integers preceding all the positive ones?
Suggestion: given a sequence in Z that includes 1 as a term, explain
why the preceding terms can’t include all the negative integers.
4. Suppose that m and n are nonnegative integers such that m
2
= 2n
2
.
Use simple facts about odd and even integers to show that m and n
6 CHAPTER I THE REAL NUMBER SYSTEM
are both even and that m/2 and n/2 are also nonnegative integers
with (m/2)
2
= 2 (n/2)
2
. Why does this imply that m = n = 0?
2 INTERVALS
Many of the features of N, Z, Q, and R are described in terms of the
numerical order of their elements. Consequently, we often work with sets
of numbers deﬁned in terms of numerical order; the simplest such sets are
the intervals. A nonempty set of numbers is called an interval if it has
the property that every number lying between elements of the set must
also be an element. We’re primarily interested in working with sets of real
numbers, so when we say a nonempty set I ⊂ R is an interval, it means
that we can prove a given x ∈ R is an element of I by simply ﬁnding
a, b ∈ I with a < x < b. However, I may well contain numbers less than
a or greater than b, so to show that a second given real number x
is also
in I we might well need to ﬁnd a different pair of numbers a
, b
∈ I with
a
< x
< b
.
We often identify intervals in terms of their endpoints. The sets
(c, d) = {x ∈ R : c < x < d} and [c, d] = {x ∈ R : c ≤ x ≤ d}
are familiar examples. The custom in the United States is to use a square
bracket to indicate that the endpoint is included in the interval and a paren
thesis to indicate that it isn’t. We should remember that intervals can also
be speciﬁed in many ways that do not involve identifying their endpoints;
our deﬁnition doesn’t even require that intervals have endpoints.
Instead of simply agreeing that (c, d) and [c, d] are intervals because
that’s what we’ve always called them, we should see that they really do
satisfy the deﬁnition. That’s the way mathematics is done. In the case
of (c, d), we should assume only that a, b ∈ (c, d) and a < x < b, and
then ﬁnd a reason why x must also be in (c, d). The transitive law for
inequalities provides all the justiﬁcation we need: for a, b ∈ (c, d) we must
have c < a and b < d, and then a < x < b implies that c < x < d. Similar
considerations explain why [c, d] is an interval.
Somewhat surprisingly, a set consisting of a single real number is an
interval. When I has a single element it is nonempty. Since we can’t
possibly ﬁnd numbers a, b ∈ I with a < b we need not worry whether
every x between elements of I satisﬁes x ∈ I. Such an interval, called
a degenerate interval, is by no means typical. Note that [c, c] always
represents a degenerate interval, but (c, c) does not represent an interval
since it has no elements.
2. INTERVALS 7
We say that an interval is closed on the right if it contains a greatest
element and open on the right if it doesn’t. We also say that an interval
is closed on the left if it contains a least element and open on the left if it
doesn’t. A closed interval is an interval that is both closed on the right and
closed on the left. It’s easy to see that every closed interval must have the
form [c, d] with c its least element and d its greatest. An open interval is
an interval that is both open on the right and open on the left. While every
interval of the form (c, d) is an open interval, there are other possibilities
to consider. We’ll return to this point later in this section.
We say that an interval I is ﬁnite (or bounded) if there is a number
M such that every x ∈ I satisﬁes x ≤ M. Every closed interval is ﬁnite
because every x ∈ [c, d] satisﬁes x ≤ c + d. Open intervals may be
either ﬁnite or inﬁnite. The set P of all positive real numbers is an example
of an inﬁnite open interval. With inﬁnite intervals it’s convenient to use the
symbol −∞ or ∞ in place of an endpoint; for example, we write (0, ∞)
for P. We don’t use a square bracket next to −∞ or ∞ because these
symbols do not represent elements of any set of real numbers.
We often use intervals to describe the location of a real number that
we only know approximately. The shorter the interval, the more accurate
the speciﬁcation. While it can be very difﬁcult to determine whether two
real numbers are exactly equal or just very close together, mathematicians
generally assume that such decisions can be made correctly; that’s one of
the basic principles underlying calculus. In fact, we assume that it’s always
possible to ﬁnd an open interval that separates two given unequal numbers.
We’ll take that as an axiom about the real numbers, rather than search for
some other principle that implies it.
Axiom 1: Given any two real numbers a and b, either a = b or there
is an ε > 0 such that a / ∈ (b −ε, b +ε).
Of course, any ε between 0 and b −a should work when a = b;
one of the things the axiom says is that b −a is positive when a = b.
In particular, two different real numbers can never be thought of as being
arbitrarily close to each other. That’s why we say, for example, that the
repeating decimal 0.9 and the integer 1 are equal; there is no positive
distance between them. We often use the axiom to prove that two real
numbers a and b are equal by proving that a ∈ (b −ε, b +ε) for every
positive ε.
While a single interval may represent an inexact speciﬁcation of a
real number, we often use sequences of intervals to specify real numbers
exactly. For example, it is convenient to think of a nonterminating decimal
8 CHAPTER I THE REAL NUMBER SYSTEM
as specifying a real number this way. When we read through the ﬁrst n
digits to the right of the decimal point and ignore the remaining ones, we’re
specifying a closed interval I
n
of length 10
−n
. For example, saying that
the decimal expansion for π begins with 3.1415 is equivalent to saying
that π is in the interval [3.1415, 3.1416]. The endpoints of this interval
correspond to the possibilities that the ignored digits are all 0 or all 9.
The complete decimal expansion of π would effectively specify an inﬁnite
sequence of intervals, with π the one real number in their intersection,
the mathematical name for the set of numbers common to all of them.
With {I
n
}
∞
n=1
being the sequence of closed intervals corresponding to the
complete decimal expansion of π, we write
∞
n=1
I
n
= {π} ;
that is, the intersection of all the intervals I
n
in the sequence is the set
whose only element is π.
In using the intersection of a sequence of intervals to deﬁne a real
number with some special property, there are two things we have to check.
The intersection can’t be empty, and every other real number except the one
we’re deﬁning must be excluded. There are some subtleties in checking
these conditions, so to simplify the situation we usually try to work with
sequences {I
n
}
∞
n=1
such that each interval I
n
in the sequence includes the
next interval I
n+1
as a subinterval. We call such a sequence a nested
sequence of intervals; the key property is that I
n+1
⊂ I
n
for all n ∈ N.
For any nested sequence of intervals,
m
n=1
I
n
= I
m
for all m ∈ N,
so we can at least be sure that every ﬁnite subcollection of the intervals in
the sequence will have a nonempty intersection. However,
∞
n=1
I
n
can
easily be empty, even for nested intervals; deﬁning I
n
= (n, ∞) provides
an easy example. We can rule out such simple examples if we restrict our
attention to closed intervals. By using any of the common statements of the
completeness property it is possible to showthat the intersection of a nested
sequence of closed intervals can’t be the empty set. But instead of proving
this as a theorem, we’ll take it as an axiom; it is easier to understand than
the assumptions we would need to make to prove it. It’s also fairly easy to
work with.
2. INTERVALS 9
Axiom 2: Every nested sequence of closed intervals in R has a
nonempty intersection.
Now let’s look at the second part of the problem of using a sequence
of intervals to deﬁne a real number. How can we know that the sequence
excludes every other real number except the one we’re trying to deﬁne?
Thanks to Axiom1, that’s easy. Let’s say {[a
n
, b
n
]}
∞
n=1
is a nested sequence
of closed intervals, with r a point in
∞
n=1
[a
n
, b
n
]. Given any x = r, we
know there is an ε > 0 such that x / ∈ (r −ε, r +ε), and then x can’t be
in any interval that includes r but has length less than ε. So our nested
sequence {[a
n
, b
n
]}
∞
n=1
can be used to deﬁne a real number if the sequence
has the property that for each ε > 0 there is an m with b
m
− a
m
< ε.
We’ll indicate this condition by writing b
n
− a
n
→ 0, anticipating the
notation of Chapter 3 for the limit of a sequence. Note that when we write
b
n
−a
n
→ 0, we are indicating a property of the sequence as a whole, not a
property of a particular term in the sequence. It’s also helpful to recognize
that for a nested sequence of closed intervals,
b
n
−a
n
≤ b
m
−a
m
for all n ≥ m.
Earlier we explained how a nonterminating decimal expansion could
be interpreted as a sequence of intervals. Note that the intervals form a
nested sequence of closed intervals. The length of the nth interval in the
sequence is 10
−n
, and obviously 10
−n
→ 0. After all, we can use the
decimal expansion of any given ε > 0 to ﬁnd n with 10
−n
< ε. So we see
that our axioms guarantee that every nonterminating decimal expansion
deﬁnes exactly one real number.
Here is a more surprising consequence of Axiom 2: it is impossible to
write a sequence of real numbers such that every real number in a closed
nondegenerate interval appears as a term. This fact about the real number
system was discovered by the German mathematician Georg Cantor in
the nineteenth century. He assumed that a sequence of real numbers had
been given, and then used the given sequence to produce a real number
that could not have been included. This is easy to do inductively. First
choose a closed nondegenerate interval that excludes the ﬁrst number in
the sequence, then successively choose closed nondegenerate subintervals
that exclude the successive numbers in the sequence. Then any number in
all the intervals can’t be in the sequence.
We often locate real numbers by the bisection method, using Axiom
2 as the theoretical basis. We start with a closed interval and successively
take either the closed right half or the closed left half of the interval just
selected. That is, if I
n
= [a
n
, b
n
], then the next interval I
n+1
is [a
n
, m
n
] or
10 CHAPTER I THE REAL NUMBER SYSTEM
[m
n
, b
n
], where m
n
is the midpoint of I
n
. This always makes b
n
−a
n
→ 0,
so however we decide which halves to choose we will wind up with exactly
one real number common to all the intervals.
We’ll conclude this section by using the bisection method to prove a
theorem describing all open intervals.
Theorem 2.1: Every ﬁnite open interval has the form (a, b) for
some pair of real numbers a, b with a < b. Every other open interval has
one of the forms (−∞, ∞), (a, ∞), or (−∞, b).
Proof : Let I be an open interval. Since our deﬁnition of interval
requires that I = ∅, we may assume that we have a real number c ∈ I.
Since I is open, we know that c is neither the least nor the greatest number
in I. We’ll examine the portions of I to the right and left of c separately,
beginning with the right.
Either (c, ∞) ⊂ I or there is a real number d > c with d / ∈ I. In the
latter case, we prove that there is a real number b > c with both (c, b) ⊂ I
and I ∩ [b, ∞) = ∅, making b the right endpoint for I. To produce b,
we let I
1
= [c, d], and deﬁne a nested sequence of closed intervals by the
bisection method. Our rule is simple: when the midpoint of I
n
is in the
original interval I, we deﬁne I
n+1
to be the closed right half of I
n
. When
the midpoint isn’t in I, we deﬁne I
n+1
to be the closed left half of I
n
.
Then each interval in our nested sequence has its left endpoint in I and its
right endpoint outside I. It’s important to note that since I is an interval
we know that every point between c and any one of the left endpoints is in
I, while no point to the right of any one of the right endpoints can be in I.
Now we deﬁne b to be the one real number common to all the closed
intervals in the nested sequence just constructed. Since the ﬁrst interval
was [c, d], we know that b ≥ c. Note that if x > b then x can’t be in I
because when the length of I
n
is less than x −b we know x must be to the
right of the right endpoint of I
n
. Since I has no greatest element, we can
also conclude that b / ∈ I and so I ∩[b, ∞) = ∅. Since c ∈ I, we must have
c < b. Our next task is to show (c, b) ⊂ I. Given any x in (c, b), we know
x can’t be in I
n
when the length of I
n
is less than b −x, so x must be to the
left of the left endpoint of I
n
. That puts x between c and the left endpoint
of I
n
, so x ∈ I and we’ve proved that (c, b) ⊂ I.
Now we consider numbers below c. Either (−∞, c) ⊂ I or there is
a number e < c with e / ∈ I. In the latter case, we can argue as above to
produce a real number a < c with (a, c) ⊂ I but I ∩ (−∞, a] = ∅.
When we sort through the various possibilities, we conclude that ev
ery open interval must have one of the four forms given in the statement
of the theorem. For example, if (c, ∞) ⊂ I and also (a, c) ⊂ I but
EXERCISES 11
I ∩ (−∞, a] = ∅, then we can conclude that I = (a, ∞). We ﬁnish the
proof by noting that (a, b) is the only one of these four forms that represents
a ﬁnite interval.
Theorem 2.1 is actually a deep result about R; it’s another version of
the completeness principle. See Exercise 9 below for a brief development
of this idea. One of the recurring problems we’ll encounter is to produce a
real number with some special property, and one of the ways to do so is to
identify it as an endpoint of a carefully deﬁned open interval. However, in
the next section we’ll develop procedures that are somewhat easier to use.
EXERCISES
5. Use the deﬁnition of interval to prove that if I and J are intervals
with I ∩ J = ∅, then both I ∩ J and I ∪ J are intervals.
6. For I
n
=
_
1
n
,
2
n
¸
, show that
∞
n=1
I
n
= ∅. What hypothesis in the
nested interval axiom does this sequence fail to satisfy?
7. For I
n
=
_
0,
1
n
_
, show that
∞
n=1
I
n
= ∅. What hypothesis in the
nested interval axiom does this sequence fail to satisfy?
8. Let {d
n
}
∞
n=1
be the sequence in {0, 1} deﬁned by d
n
= 1 when n is
a perfect square and d
n
= 0 when n is not a perfect square. Prove
that 0.d
1
d
2
d
3
. . . d
n
. . . is a nonterminating, nonrepeating decimal.
9. For {[a
n
, b
n
]}
∞
n=1
an arbitrary nested sequence of closed intervals,
show that
∞
n=1
(−∞, a
n
) is an open interval, and prove that its right
endpoint must be an element of
∞
n=1
[a
n
, b
n
].
10. Use the bisection method to prove that 2 has a positive square root in
R. The crucial fact is that if 0 < x < y, then x
2
< y
2
. Begin your
search for
√
2 in the interval [1, 2].
11. Without knowing that
√
2 exists, how could you deﬁne an open inter
val whose left endpoint is
√
2?
3 SUPREMA AND INFIMA
Working with intervals that fail to have a least or greatest element is usually
easy, since an endpoint of the interval can be an adequate substitute. For
sets of real numbers that are not intervals, it’s often useful to have numbers
that play the role of endpoints. In place of the left endpoint of E, we
might use the greatest a such that E ⊂ [a, ∞), and in place of the right
endpoint, we might use the least b such that E ⊂ (−∞, b]. As the case
of open intervals illustrates, such numbers aren’t necessarily elements of
the given set, so we can’t quite call them least or greatest elements, but
12 CHAPTER I THE REAL NUMBER SYSTEM
that’s almost the concept we have in mind. To suggest that notion without
quite saying it, we make use of the Latin words inﬁmum and supremum,
meaning lowest and highest, and indicate them symbolically as inf E and
supE. The familiar English words inferior and superior come from the
same Latin roots, so the meanings of these terms are easy to keep straight.
Here’s a formal deﬁnition of these terms as we will use them.
Definition 3.1: For E a nonempty set of real numbers, we say that
the real number b is the supremum of E, and write b = supE, if every
x ∈ E satisﬁes x ≤ b but for each c < b there is an x ∈ E with x > c.
Similarly, we say that the real number a is the inﬁmum of E, and write
a = inf E, if every x ∈ E satisﬁes x ≥ a but for each c > a there is an
x ∈ E with x < c.
This deﬁnition deserves to be examined quite carefully. Saying that
every x ∈ E satisﬁes x ≤ b is equivalent to the settheoretic inclusion
E ⊂ (−∞, b]. We describe this relationship by saying that b is an upper
bound for E. Saying that some x ∈ E satisﬁes x > c is then equivalent to
saying that c is not an upper bound for E. Thus, according to the deﬁnition,
the supremum of a set is an upper bound with the property that no smaller
number is an upper bound for the set, so the supremumof a set is sometimes
called the least upper bound. Similarly, the inﬁmum is sometimes called
the greatest lower bound. But it’s all too easy to confuse least upper
bounds with least elements or greatest lower bounds with greatest elements,
so we’ll stick with the names supremum and inﬁmum.
By the way, as we compare values of real numbers and describe the
results, we often use the words less, least, greater, or greatest rather than
smaller, smallest, larger, or largest. When we’re comparing positive quan
tities it makes little difference, and the latter terms are more common in
everyday use. But while we easily recognize that −1, 000, 000 is less than
−1, most people don’t think of it as smaller; the word smaller suggests
being closer to zero instead of being farther to the left. In general, we
should avoid using any terms that suggest the wrong concept even if we
know our use of the terms is correct; being right but misunderstood isn’t
much better than being wrong.
Returning to the notion of supremum and inﬁmum, there’s an obvious
question: why should a set of real numbers have a supremum or inﬁmum?
In fact, some hypotheses about the nature of the set are essential, but
according to the theorem below, they are remarkably simple.
3. SUPREMA AND INFIMA 13
Theorem 3.1: Every nonempty set of real numbers that has an
upper bound has a supremum, and every nonempty set of real numbers that
has a lower bound has an inﬁmum.
Some mathematicians prefer to call this theorem an axiom because
it’s another version of the completeness property for R. It can serve as a
logical substitute for our axiom about nested sequences of closed intervals;
when this theorem is assumed to be true, our axiom can be derived as
a consequence. Conversely, we can and do use our axiom to prove the
theorem. We use bisection to locate the supremum and inﬁmum with
almost the same argument we used to prove our theorem about the form of
open intervals.
Proof : Let’s begin by assuming that E is a nonempty set of
numbers, and that Ehas an upper bound b
1
. There must be a number x ∈ E,
and if we choose a number a
1
< xthen a
1
is not an upper bound for E. Thus
I
1
= [a
1
, b
1
] is a closed interval whose right endpoint is an upper bound
for E, but whose left endpoint is not. Using bisection and mathematical
induction, we can produce a nested sequence of closed intervals {I
n
}
∞
n=1
,
each interval having these same properties. For I
n
= [a
n
, b
n
], its midpoint
is m
n
=
1
2
(a
n
+b
n
). We deﬁne
I
n+1
=
_
[a
n
, m
n
], m
n
an upper bound for E
[m
n
, b
n
], m
n
not an upper bound for E.
Then each I
n
is a closed interval containing the next interval in the sequence
and the lengths satisfy b
n
−a
n
→ 0. Once again, there is exactly one real
number common to all these intervals; let’s call it s.
Now let’s see why s = sup E. For any given x > s, there is an interval
I
n
in the sequence with length less than x − s, so its right endpoint is to
the left of x. That shows that x is greater than an upper bound for E, so
x / ∈ E when x > s. Hence every x ∈ E satisﬁes x ≤ s. On the other
hand, given any y < s, there is an interval I
n
with its left endpoint to the
right of y. Since that endpoint isn’t an upper bound for E, we see that y
isn’t either; some z ∈ E satisﬁes y < z. We’ve proved that s = sup E.
The same sort of argument is used to prove that every nonempty set of
real numbers with a lower bound has an inﬁmum. We leave that proof to
the reader.
Our ﬁrst use of the existence of suprema and inﬁma will be to establish
a useful substitute for Axiom 2. There are times when it’s convenient to
deﬁne a real number as the one point common to all the intervals in some
collection when the collection is not a nested sequence. The theorembelow
14 CHAPTER I THE REAL NUMBER SYSTEM
is a real help in such situations. In particular, we’ll use it in Chapter 5 to
help develop the theory of integration.
Theorem 3.2: Let I be a nonempty family of closed intervals such
that every pair of intervals in I has a nonempty intersection. Then there is
at least one point that is common to every interval in I.
Proof : For Lthe set of left endpoints of intervals in I, we’ll prove
that L has a supremum and that supL is in every interval in I. We’re given
that I is a nonempty family of closed intervals; let’s consider an arbitrary
[a, b] ∈ I. Clearly a ∈ L, and if we can prove that b is an upper bound for
L then we’ll know that L has a supremum. We’ll also know that
a ≤ supL ≤ b.
Given any other interval [c, d] ∈ I, its intersection with [a, b] is
nonempty. So there must be an x ∈ Rwith
x ∈ [a, b] ∩ [c, d].
That means both
a ≤ x ≤ b and c ≤ x ≤ d,
so
c ≤ x ≤ b.
Thus c ≤ b for every [c, d] ∈ I, so b is an upper bound for L and therefore
supL ∈ [a, b]. Since [a, b] represents an arbitrary interval in I, this proves
that supL is in every interval in I.
Another useful notion closely related to the supremum and inﬁmum
of a set is the set’s diameter. Geometers use the word diameter to refer
to a line segment that joins two points on a circle and passes through the
center of the circle. Of all the line segments joining two points on a given
circle, the diameters have the greatest length; that length is what most of
us mean by the diameter of a circle. Sets of real numbers don’t look very
much like circles, but we can still talk about the length of a line segment
connecting two numbers in the set. Of course, we calculate that length by
subtracting the smaller number from the larger, so the set of all lengths of
line segments linking two points in the set E of real numbers is
L = {x −y : x, y ∈ E}.
3. SUPREMA AND INFIMA 15
If L has a supremum, we call that supremum the diameter of E. It’s not
too hard to see that the diameter of E can be deﬁned if and only if E is
a nonempty set that has both an upper and a lower bound. The diameter
of a closed interval is simply its length, and this is one of the cases where
the set of lengths has a largest element. The theorem below extends this
formula for the diameter to more general sets.
Theorem 3.3: Let E be a nonempty subset of R. If E has both an
upper and a lower bound, then
diamE = sup E −inf E.
Proof : As with most formulas involving suprema and inﬁma,
several steps are required to verify it. We need to show that supE − inf E
satisﬁes the deﬁnition of supL. First, we note that E ⊂ [inf E, supE], so
for every possible choice of x, y ∈ E we have
x −y ≤ supE −inf E,
proving that supE − inf E is an upper bound for L. That’s the ﬁrst step.
To complete the proof, we show that for every b < supE − inf E, there
is an element of L that is greater than b.
Given b < supE −inf E, we call
ε = sup E −inf E −b.
Then ε > 0, so
supE −
ε
2
< supE and inf E < inf E +
ε
2
.
Then by the deﬁnition of supE and inf E, there are x, y ∈ E such that
x > supE −
ε
2
and y < inf E +
ε
2
.
These elements of E satisfy
x −y ≥ x −y > supE −inf E −ε,
and the quantity on the right is just b.
In many problems we ﬁnd that we need the supremum or inﬁmum of a
set of sums of numbers, with each term in the sum representing an element
of some simpler set. As the theorem below indicates, we can deal with
such problems by analyzing the component sets one at a time.
16 CHAPTER I THE REAL NUMBER SYSTEM
Theorem 3.4: Suppose that {E
1
, E
2
, . . . , E
n
} is a collection of sets
of numbers and that each of the sets has a supremum. Then
sup
_
n
k=1
x
k
: x
1
∈ E
1
, x
2
∈ E
2
, . . . , x
n
∈ E
n
_
=
n
k=1
supE
k
.
Similarly, if each of the sets E
k
has an inﬁmum, then the inﬁmum of the set
of sums is the sum of the inﬁma.
Proof : Once we understand the case n = 2, it’s a simple matter
to complete the proof by using induction. For the inductive step, we
note that when our collection of sets is expanded from {E
1
, . . . , E
n
} to
{E
1
, . . . , E
n
, E
n+1
}, we can regroup
n+1
k=1
x
k
as (
n
k=1
x
k
)+x
n+1
, with
each of these two numbers representing an element of a set with a known
supremum. With that in mind, let’s simply assume that E and F are sets
of numbers and that both sets have suprema. Then we show that
sup{x +y : x ∈ E and y ∈ F} = sup E + supF.
By assumption, neither E nor F can be the empty set, so
S = {x +y : x ∈ E and y ∈ F}
is not the empty set. Every x ∈ E satisﬁes x ≤ supE and every y ∈ F
satisﬁes y ≤ supF, so every x +y ∈ S must satisfy
x +y ≤ supE + supF.
On the other hand, given any c < supE + supF, if we call
ε = supE + supF −c,
then there must be an x
c
∈ E with x
c
> supE −ε/2 and there must be a
y
c
∈ F with y
c
> supF −ε/2. So we have
x
c
+y
c
∈ S and x
c
+y
c
> supE + supF −ε = c,
proving that supE + supF is the supremum of S.
Mathematicians often introduce sets without having a very good idea
how to determine what all the elements are, just as they use letters to rep
resent numbers whose precise values they may not know how to compute.
We should think of supE, inf E, and diamE in much the same spirit.
Determining their exact values requires considering the numbers in a set
all at once, not just one at a time, and for complicated sets that may be
EXERCISES 17
harder than anything we know how to do. That’s good news for math
ematicians; the difﬁculty in ﬁnding these values makes the symbols for
them more useful. When we use these symbols to formulate and prove
relationships, they help us organize complex systems into simpler patterns,
neatly avoiding the complexity of the components. That’s the sort of thing
that makes abstract mathematics a practical skill, not just an intellectual
exercise.
EXERCISES
12. Complete the proof that if E is a nonempty set of real numbers and
E has a lower bound, then E has an inﬁmum.
13. Provide an argument showing that if E and F are sets of real numbers
and each has an inﬁmum, then
inf {x +y : x ∈ E and y ∈ F} = inf E + inf F.
14. Theorem 2.1 can also be used to prove the existence of suprema; this
exercise shows how. Suppose that E ⊂ R, and let B be the set of all
upper bounds of E. Without assuming that E has a supremum, show
that
U =
_
b∈B
(b, ∞)
deﬁnes an open interval as long as B = ∅. Then show that when
E = ∅, the left endpoint of U satisﬁes the deﬁnition of supE.
15. When E ⊂ R, it is convenient to deﬁne the set
−E = {−x : x ∈ E} = {y : −y ∈ E}.
Prove that E has a supremum if and only if −E has an inﬁmum
and that inf (−E) = −supE. Note: The second part of a statement
like this is generally assumed to include the ﬁrst part implicitly; the
identity of two expressions requires both to be deﬁned in the same
cases. Exceptions should be noted explicitly.
16. Find the diameter of the set
_
1
n
: n ∈ N
_
.
4 EXACT ARITHMETIC IN R
Exact arithmetic is standard operating procedure in the rational number
system Q; the rules are
m
n
+
m
∗
n
∗
=
mn
∗
+nm
∗
nn
∗
,
m
n
·
m
∗
n
∗
=
mm
∗
nn
∗
, and
m
n
_
m
∗
n
∗
=
mn
∗
nm
∗
18 CHAPTER I THE REAL NUMBER SYSTEM
Consequently, the familiar rules of algebra implicitly assume that the in
dicated arithmetic operations are performed exactly. Of course, when we
do arithmetic using decimals, we often wind up with approximate values
instead of exact values. If we’re not careful that can cause some problems;
even the best computers are subject to errors introduced when decimal val
ues are rounded off. Here’s a simple example of the sort commonly used
to introduce calculus operations.
Example. Compute representative values of
(1 +h)
2
−1
h
for h in a small
interval (0, ε).
In terms of exact arithmetic, the fraction is exactly equal to 2 + h, so
the resulting values are all close to 2. But if we ask a computer to evaluate
the original expression with h = 10
−50
, the computer is likely to return a
value of 0. While most computers are capable of recognizing 10
−50
= 0,
they ﬁnd it difﬁcult to tell that
_
1 + 10
−50
_
2
= 1 because in processing
numerical values digitally they usually ignore digits too far to the right of
the ﬁrst nonzero one. Consequently, the computer will treat the original
fraction as 0 divided by a nonzero number.
For arithmetic operations in the real number system R, the usual rules
of algebra are valid, assuming that we do exact arithmetic. But if a number
is known to us only as a description of a process for forming a nested
sequence of closed intervals, how are we supposed to do arithmetic with
it? The way we answer such a question depends on what we understand a
real number to be. The fact is that real numbers may only be understood
as approximations, and given any approximation we can imagine a context
in which that approximation won’t be adequate.
Mathematicians are professional pessimists. We tend to believe that
anything that can possibly go wrong certainly will go wrong sooner or
later, probably sooner than expected, and with worse consequences than
previously imagined. So to specify a real number when the exact value can’t
be conveniently determined, we don’t settle for a single approximation; we
want a whole family of approximations that includes at least one for every
purpose. Then when we’re told how accurate the approximation needs to
be, we can select one that meets the requirements. A nested sequence of
closed intervals that includes arbitrarily short intervals is a good example
since we can regard it as a family of approximations to the one number a
common to all the intervals. When we need an approximation x to a that
is accurate to within ε, we simply ﬁnd one of the intervals that is shorter
than ε and choose any x from that interval.
The philosophy in performing arithmetic on real numbers is that ap
proximations to the input can be used to produce approximations to the
4. EXACT ARITHMETIC IN R 19
output. Let’s examine how it works on the fundamental arithmetic opera
tions of addition, multiplication, and division. Subtraction is so much like
addition that there’s no need to treat it separately. The idea is to assume
that x and y represent approximations to the real numbers a and b, perform
the operation under consideration on x and y as well as on a and b, and
then express the difference of the results in terms of x − a and y − b. We
then analyze this expression and ﬁgure out how small x −a and y −b
need to be.
Addition is especially simple. We see that
(x +y) −(a +b) = (x −a) + (y −b)
≤ x −a +y −b.
To obtain (x +y) −(a +b) < ε, we require that both x −a < δ and
y −b < δ with δ =
1
2
ε.
Multiplication is more complicated. Since
xy −ab = (x −a) (y −b) +a (y −b) +b (x −a),
we see that
xy −ab ≤ x −a y −b +a y −b +b x −a.
If we require both x −a < δ and y −b < δ, we have
xy −ab < δ
2
+a δ +b δ = δ (δ +a +b),
so we choose δ > 0 to make this last product be less than ε. Its second
factor is bounded by 1 + a + b when δ < 1, so we can guarantee that
xy −ab < ε by requiring x −a < δ and y −b < δ for any δ ∈ (0, 1)
that satisﬁes
δ ≤
ε
1 +a +b
.
In particular, we can always use
δ =
ε
ε + 1 +a +b
to satisfy our requirements.
To approximate a/b by x/y, we ﬁrst must assume that b = 0. Then
we restrict our attention to y with y −b < b to guarantee that y = 0.
We then ﬁnd that
x
y
−
a
b
=
bx −ay
by
=
b (x −a) −a (y −b)
by
. (1.1)
20 CHAPTER I THE REAL NUMBER SYSTEM
If we require both x −a < δ and y −b < δ, then we have a simple
bound for the numerator in (1.1):
b (x −a) −a (y −b) ≤ b x −a +a y −b < (b +a) δ.
If we also require 0 < δ < b, then the denominator in (1.1) satisﬁes
by = b y > b (b −δ).
Hence taking absolute values in (1.1) gives the bound
¸
¸
¸
¸
x
y
−
a
b
¸
¸
¸
¸
=
b (x −a) −a (y −b)
by
<
(b +a) δ
b (b −δ)
.
Now we choose δ in (0, b) to make sure that this last quantity is at most
ε. For b −δ > 0, the inequality
(b +a) δ
b (b −δ)
≤ ε
is satisﬁed when
(b +a) δ ≤ b (b −δ) ε = b
2
ε −b δε,
or, equivalently, when
(a +b +b ε) δ ≤ b
2
ε.
Consequently, as long as b = 0 we can choose
δ =
b
2
ε
a +b +b ε
;
this also gives us that 0 < δ < b. Then requiring x −a < δ and
y −b < δ will guarantee that x/y −a/b < ε, for whatever ε > 0 is
given.
It’s important to note the form of our answers. For any given ε > 0, we
guarantee that the approximate answer is within ε of the exact answer as
long the approximate input is within δ of the exact input. The quantitative
relationship between δ and ε changes as the arithmetic operation changes,
but the logical relationship stays the same. That, in fact, is why we prefer
to use the name δ and not the formula we ﬁnd for it. With just a little
imagination, we can then recognize that the same sort of conditions can be
developed for more complicated arithmetic operations. The quantitative
relationship between δ and ε can be built up in stages, just as the operation
EXERCISES 21
is built in stages using the basic building blocks of addition, multiplication,
and division.
Here’s a simple example showing how it’s done. Let’s say x, y, and
z are approximations to a, b, and c, respectively, and we’re going to use
x(y +z) as an approximation to a (b +c). We assume that ε > 0 has been
given. We need a δ > 0 such that whenever x −a, y −b, and z −c are
all smaller than δ, we can be certain that x(y +z) is within ε of a (b +c).
The last thing we do in calculating x(y +z) is to multiply x and y + z,
and we know enough about approximate multiplication to know there is a
δ
1
> 0 such that
x(y +z) −a (b +c) < ε
as long as
x −a < δ
1
and (y +z) −(a +b) < δ
1
.
Now we use δ
1
as another value for ε; our knowledge of approximate
addition tells us there is a δ
2
> 0 such that
(y +z) −(a +b) < δ
1
as long as
y −b < δ
2
and z −c < δ
2
.
We can therefore choose δ > 0 to be the smaller of δ
1
and δ
2
, which is δ
2
in this case.
That’s how we interpret exact arithmetic when only approximations
are possible. We can guarantee that the output possesses whatever level
of accuracy is desired by simply making sure that the input has sufﬁcient
accuracy; using exact input is not crucial.
EXERCISES
17. Given a, b with a +b > 0, explain why there must a δ > 0 such that
u +v > 0 for all u, v with u −a < δ and v −b < δ.
18. Given a, b, c, d with cd = 0, explain why whenever ε > 0 is given,
there must be a δ > 0 with the property that every x, y, z, and w with
x −a, y −b, z −c, and w −d all less than δ must satisfy
¸
¸
¸
¸
xy
zw
−
ab
cd
¸
¸
¸
¸
< ε.
22 CHAPTER I THE REAL NUMBER SYSTEM
19. For x an approximation to a and a = 0, the quantity (x −a) /a is
called the relative error. When x and y are approximations to a and
b and xy is used to approximate ab, show how to express the relative
error of the product in terms of the relative errors of the factors.
20. When a = 0 is approximated by x, we use 1/x to approximate 1/a.
Express the relative error in the latter approximation in terms of the
relative error in the ﬁrst.
21. For x ∈ [a, b] and y ∈ [c, d], it’s always true that x+y ∈ [a +c, b +d]
and x − y ∈ [a −c, b −d]. Is it always true that xy ∈ [ac, bd]? Or
that x/y ∈ [a/c, b/d]?
5 TOPICS FOR FURTHER STUDY
One method that mathematicians have used to describe the real number
system is to identify real numbers with splittings of the rational numbers
into two nonempty complementary sets, one of which has no greatest
element but contains every rational number less than any of its elements.
Such splittings are called Dedekindcuts. In this scheme, every real number
α is identiﬁed with a set of rationals that we would think of as having α
for its supremum. All the properties of the real number system can be
developed from this representation of real numbers; a complete exposition
of this theory can be found in the classical text of Landau [2].
Open and closed intervals are the simplest examples of open and closed
sets of numbers, two fundamental notions in topology. An open set of
numbers is a set that contains an open interval about each of its elements.
Every open set can be written as the union of a sequence of pairwise disjoint
open intervals. Closed sets of numbers are characterized by the property of
containing the supremum and inﬁmum of each of their bounded, nonempty
subsets. A closed set can also be described as a set whose complement
is open. Closed sets can be much more complex than open sets, and
there is no way to characterize them in terms of unions of closed intervals.
For example, the Cantor set is a closed subset of [0, 1] that contains no
nondegenerate intervals, but it has so many points in it that no sequence of
real numbers can possibly include all of them.
II
CONTINUOUS FUNCTIONS
A
s we try to make sense of our surroundings, we ﬁnd that evolu
tionary change seems much simpler than abrupt change, because
it is easier to recognize the underlying order. The same is true in
mathematics; we need continuity to recognize order in potentially chaotic
models.
1 FUNCTIONS IN MATHEMATICS
It seems that every modern calculus book discusses functions somewhere
in the ﬁrst chapter, often expecting the concept to be familiar from precal
culus courses. Since functions are so commonplace, it’s remarkable that
our concept of function is a fairly recent one, unknown to the founders of
calculus and quite different from what was expected by the ﬁrst mathe
maticians to use abstract functions. The original idea was simply to extend
algebraic notation in a way that lets us work with quantitative relationships
abstractly, in much the same way that variables make it possible to look
23
24 CHAPTER II CONTINUOUS FUNCTIONS
at arithmetic abstractly and develop general rules. Just as we often treat
variables as unknowns and try to solve for them, we also sometimes treat
functions as unknowns, and many important problems involve our trying
to solve for them. But some of the things people found when they went
looking for functions turned out to be very different from what they ex
pected to ﬁnd, and the need for careful deﬁnitions became apparent. Now
we use functions to describe just about everything that can be considered a
mathematical operation.
Mathematical operations don’t just happen all by themselves; some
thing must be operated on. And the operation must result in something,
even if that something is a concept we associate with nothing, such as zero
or the empty set. So we naturally associate mathematical operations with
two sets, one for the input and one for the output. (Here we’re using the no
tion of set in its most general sense—elements of the input or output might
themselves be sets of numbers, for example.) The operation is called a
function if the result of performing it depends only on the designated input,
with a special emphasis on the word designated. For example, if the set
for the input is a set of coins and the operation to be performed on a coin
consists of determining its denomination, that’s a function. However, if the
operation consists of ﬂipping the coin and noting whether it lands heads or
tails, that’s not a function because the outcome clearly depends on some
thing besides the coin that was ﬂipped. With this example in mind, here’s
a formal deﬁnition that spells out just when we can think of an operation
as a function.
Definition 1.1: Let A and B be sets, and let f be any process that
can accept each element x of A as input and then produce an element of
B as output. If the element in B that results is entirely determined by the
element x selected from A, we say that f is a function from A to B. We
indicate this symbolically as f : A → B and write f (x) for the result of
performing the operation f on the element x. We call the set Athe domain
of the function and the set B the range of the function. The subset of B
given by {f (x) : x ∈ A} is called the image of A under f.
The general concept of function is often illustrated concretely by ex
amples where the domain or range consists of physical objects, such as the
function whose domain is the set of all horses present in a given region
at a given time and whose range is the set of integers, with the value of
the function at each horse being its age in years. However, such exam
ples fail to convey an important aspect of the way functions are used in
mathematics. When the domain and range of a function are composed of
1. FUNCTIONS IN MATHEMATICS 25
mathematical objects (such as numbers) rather than physical objects, the
rules change signiﬁcantly. Physical objects often are one of a kind with a
transitory existence; for example, if no two pieces of charcoal are exactly
alike, then after you burn one to measure its value as an energy source you’ll
never be able to burn the same piece again. By virtue of their abstractness,
mathematical objects are not subject to the same limitations. When you
add 1 to 3 and get an answer of 4, you don’t change 3 at all; the concept
of 3 remains ﬁxed. More importantly, a single mathematical object can
appear in a variety of guises. We think of 2 + 1, 6/2, and
√
9 as exactly
the same number as 3. So if f is a function and the number 3 is an element
of its domain, then f (3), f (2 + 1), f (6/2), and f
_√
9
_
must all refer the
same element of the range of f. To a mathematician, writing f : A → B
includes the requirement that when f operates on different representations
of the same element of A, the results all represent the same element of B.
An understanding about when different expressions represent the same
element of the domain or of the range is a necessary part of the deﬁnition
of any particular function, even though such understandings are rarely
stated explicitly. Such understandings explain why, for example, there is
no function f : R →R with the property that f
_
x
2
_
= x for all x ∈ R;
it’s entirely possible to have x
2
1
= x
2
2
but x
1
= x
2
. And we can’t deﬁne
g : Q →Qby g
_
m
n
_
=
m+2
n
, because the value of
m+2
n
depends on both
mand n, not just on the value of
m
n
as an element of Q. Subtle distinctions
like this are an inescapable part of modern mathematics. After all, the
notion that different things can sometimes be considered equal and that
equal inputs should lead to equal outputs is fundamental. It’s the guiding
principle we follow when solving equations in algebra, for example. To a
large extent, it’s what separates mathematics from other things we might
do with mathematical symbols. That’s why almost every operation that’s a
part of mathematics can be considered a function.
In scientiﬁc work there’s a somewhat different aspect to the problem of
what constitutes a function. For example, suppose there are two parameters
x and y that are used to describe some system that changes over time. We
may make a number of simultaneous measurements of x and y, and then try
to determine whether y can be considered a function of x. The mathematical
criterion is simply that any two occasions producing equal xvalues must
also produce equal yvalues, regardless of when those occasions take place.
Usually the scientist can’t examine all possible values of x and y, and the
challenge is to develop a theory that explains why y should be a function of
x when the observed values seem consistent with that assumption. That’s
a key part of the scientiﬁc method.
26 CHAPTER II CONTINUOUS FUNCTIONS
In formal mathematics, a function f : A → B consists of the domain
A, the range B, and the rule used to determine f (x) for arbitrary x ∈ A.
But in calculus it’s much more common to think of the function as simply
the rule used to determine f (x). For the most part, both the domain and
range of the functions we work with are sets of real numbers; we call such
functions numerical functions. Since our understanding about when two
mathematical expressions represent the same real number generally doesn’t
change with the subsets of real numbers considered, we won’t worry about
specifying the domain or range of numerical functions unless there is a
special need to do so. For instance, there is rarely any need to treat
f : R →R deﬁned by f (x) =
1
x
2
+ 1
and
˜
f : R → [0, 1] deﬁned by
˜
f (x) =
1
x
2
+ 1
as different functions, and we’re not likely to do so. When we do use the
notation f : A → B in reference to a numerical function, it’s because we
really do want to call attention to the domain and range; at some stage a
property of one or both will be important.
In abstract mathematics, the graph of a function f : A → B is the
set of all pairs of elements (x, y) with x ∈ A, y ∈ B, and y = f (x).
The deﬁnition of function guarantees that when (x
1
, y
1
) and (x
2
, y
2
) are
in the graph of a single function f and x
1
= x
2
, then y
1
= y
2
as well.
In the case of numerical functions, both A and B are sets of real numbers
and we usually identify the pair (x, y) with the point in the Cartesian plane
having those coordinates, making the graph a set of points in the plane. In
common usage, the graph of f usually refers to a pictorial representation
of this set of points in the plane instead of to the set itself. Points (x
1
, y
1
)
and (x
2
, y
2
) in the plane have x
1
= x
2
when the points are on the same
vertical line, so the graph of a numerical function has the property that no
vertical line can intersect it in more than one point. Graphs are a powerful
tool for studying numerical functions because they present a whole range
of values of a function simultaneously, rather than one value at a time. But
the things we can do with graphs are limited, and an algebraic or analytic
representation of a function is usually better suited to computations.
In addition to operating with numerical functions, we’ll operate on
them. Numerical functions are basic elements in a more abstract level
of mathematics that we use to study the relationships between numerical
EXERCISES 27
variables, where we think of the relationships themselves as variables.
We’ll end this section with a description of some of the ways we combine
functions.
Given two functions f and g with a common domain and both having
ranges in R, we can combine their values arithmetically to create new
functions on the same domain. For example, we can always deﬁne f +g,
f − g, and fg as realvalued functions on the same domain by simply
performing the indicated operation on their values at each point in E. That
is,
(f +g) (x) = f (x) +g (x),
(f −g) (x) = f (x) −g (x),
and (fg) (x) = f (x) g (x).
We can deﬁne constant multiples of f or g by the same process, and we
can also deﬁne f/g if zero isn’t a value of g.
We can also combine functions by the operation of composition, if
their domains and ranges match up appropriately. The composite function
f ◦ g is deﬁned whenever the set of values of g is a subset of the domain
of f. In that case the domain of f ◦ g is the domain of g and the range of
f ◦ g is the range of f. All these conditions are obviously what we need to
deﬁne the value f ◦ g at x to be f (g (x)). Whenever g (x) represents an
element of the domain of f, it makes perfectly good sense to perform the
operation of f on it, and it produces the value f (g (x)).
EXERCISES
1. Find a numerical function f with domain [1, ∞) such that
f
_
_
x
2
+ 1
_
= x
2
+ 2x + 3
_
x
2
+ 1 for all x ≥ 0.
2. A moving point P in the coordinate plane is sometimes located in
terms of its polar coordinates (r, θ); its Cartesian coordinates (x, y)
are then
x = r cos θ and y = r sinθ.
While x, y, and r are considered to be functions of P, θ usually is not,
because different values of θ correspond to the same point. Usually
the value we use for θ depends on the route that was followed to arrive
at P, not just the location of the point. With that in mind, explain why
28 CHAPTER II CONTINUOUS FUNCTIONS
sin 2θ is considered a function of P, but sin
1
2
θ is not. (The domain
of the function does not include the origin.)
3. Suppose that f is a numerical function deﬁned on [a, b], and that
s is a number such that f (b) − f (a) ≥ s (b −a). Use bisection
and mathematical induction to prove that there is a nested sequence
{[a
n
, b
n
]}
∞
n=1
of closed intervals such that
b
n
−a
n
= 2
−n
[b −a] and f (b
n
) −f (a
n
) ≥ s (b
n
−a
n
)
for each interval [a
n
, b
n
] in the sequence. Your proof should show
how to construct the sequence in terms of the unspeciﬁed function f.
2 CONTINUITY OF NUMERICAL FUNCTIONS
The concept of function is so general that it is virtually impossible to
visualize generic functions from R to R. The operations used to evaluate
f (x) at different values of x can be entirely independent of each other. In
principle there could be a different formula used at each x. Such functions
have little practical value; we certainly can’t use them to model physical
processes. For example, if we can only measure temperatures to the nearest
hundredth of a degree, how could we ever make use of a function of the
temperature for which the difference between f (98.6
◦
) and f (98.5999
◦
)
were signiﬁcant? Unless approximate values for x can be used to ﬁnd
approximate values for f (x), any physical system modeled by f will
appear to behave so erratically that it’s likely to be considered random
rather than deterministic.
We deal with that problemwhen we impose the condition of continuity,
a condition that we’ll describe in several stages. For f a numerical function
deﬁned on all of R, the condition is basically the same as the conditions
we developed for doing exact arithmetic in Rby means of approximations.
That is, we say that f is continuous at the real number a if for every
given ε > 0, there is a δ > 0 such that f (x) −f (a) < ε for all x
with x −a < δ. Study this statement carefully; it says a lot, and the
logic of what it says is fairly complex. It has several distinct parts, and
its meaning is wrapped up in the precise relationship between those parts.
It’s a statement about all of the values of f on each of a family of intervals
centered at a. Trying to consider all the possible values of f (x) as x
ranges over a single interval is hard enough, and we have to consider what
happens on many different intervals simultaneously. The task is simpliﬁed
considerably when we can restrict our attention to an interval where the
values of f either always increase or always decrease, but sometimes that
isn’t possible.
2. CONTINUITY OF NUMERICAL FUNCTIONS 29
Deciding whether a given function is continuous at a given point looks
like a difﬁcult task, but the assumption of continuity is an extremely pow
erful one. In developing mathematical theories, knowledge of continuity
is equivalent to possession of a machine that takes arbitrary positive ep
silons and processes them into suitable positive deltas. That’s the good
news about continuity; the bad news is that proving a given function is
continuous often amounts to assembling such a machine.
If our function is not deﬁned on all of R but only on a subset of it,
the condition we use to deﬁne continuity must be modiﬁed. For example,
when f : [a, b] → R, there won’t be any δ > 0 for which we can say
anything about the values of f (x) at every x ∈ (a −δ, a +δ). So we do
the best we can, and require only that the condition on the values of f (x)
hold for all x in the intersection of (a −δ, a +δ) and the domain of f.
Since the condition always involves f (a), we insist that a be a point in the
domain of f before we attempt to spell out further conditions. Now we’re
ready to deﬁne continuity for numerical functions.
Definition 2.1: For f a numerical function with domain E, we
say that f is continuous at the point a provided that a ∈ E and that for
each given ε > 0 there is a δ > 0 such that every x ∈ (a −δ, a +δ) ∩ E
satisﬁes f (x) −f (a) < ε. Under the same conditions, we also say that
a is a point of continuity for f. We say simply that f is a continuous
function if every point in its domain is a point of continuity.
The beauty of this deﬁnition is that it is so thoroughly compatible
with the things we’ve done so far and the things we’ll want to do, once
we get around to them. Obviously it’s closely connected to our notion of
approximation; if a numerical function f is continuous at a point a in its
domain and we have a family of approximations to a involving only points
where f is deﬁned, then applying f to them produces a corresponding
family of approximations to f (a). It’s often useful to reverse the process
and think of f (a) as approximating all the values of f (x) at nearby points.
When f is continuous, we can say a great deal about the properties of
f (x) over an entire interval in terms of its values at relatively few well
placed points. Arbitrary functions can be wild; by comparison, continuous
functions are quite tame.
At the end of the last section, we mentioned some operations we used
to combine functions. Such operations always preserve continuity; that’s
an extremely useful fact. It’s so important that we’ll state it as two formal
theorems, even though the proofs are almost obvious.
30 CHAPTER II CONTINUOUS FUNCTIONS
Theorem 2.1: Let f andg be givennumerical functions withdomain
E. Then each point of continuity for both f and g is a point of continuity
for f +g, f −g, and fg. The same is true for f/g when zero is not in the
image of E under g.
Proof : Note that continuity appears both in the hypotheses and
in the conclusion. Effectively, the hypotheses give us some ε to δ machines
that we can work with, and we link them together to build the machine
called for in the conclusion. Our knowledge of approximate arithmetic
lets us treat all three cases with the same argument. Consider the prod
uct fg, for example. Given a ∈ E and ε > 0, we can guarantee that
(fg) (x) −(fg) (a) < ε by making sure that both f (x) −f (a) < δ
0
and g (x) −g (a) < δ
0
for some suitably chosen δ
0
> 0. Since a is a
point of continuity for f, we can choose δ
1
> 0 such that
f (x) −f (a) < δ
0
for all x ∈ (a −δ
1
, a +δ
1
) ∩ E.
Since a is also a point of continuity for g, we can also choose δ
2
> 0 such
that
g (x) −g (a) < δ
0
for all x ∈ (a −δ
2
, a +δ
2
) ∩ E.
So we choose δ to be the smaller of the two positive numbers δ
1
and δ
2
;
then every x ∈ (a −δ, a +δ) ∩E satisﬁes both conditions. Consequently,
every x ∈ (a −δ, a +δ)∩E satisﬁes (fg) (x) −(fg) (a) < ε, and we’ve
proved continuity at a.
Theorem 2.2: Let f and g be numerical functions, with the range of
g contained in the domain of f. If g is continuous at c and f is continuous
at g (c), then f ◦ g is continuous at c.
Proof : As long as we understand what everything means, the
proof of this theorem is straightforward. We should begin by noting that
under the hypotheses, f ◦ g is a numerical function with the same domain
as g. Let’s call this set E and call b = f (c). Thus we need to show that for
any given ε > 0, there is a δ > 0 such that every x ∈ (c −δ, c +δ) ∩ E
satisﬁes
f (g (x)) −f (b) < ε.
Since we’ve assumed that f is continuous at b, we know that for our given
ε there is a δ
0
> 0 such that every y in the intersection of (b −δ
0
, b +δ
0
)
and the domain of f satisﬁes f (y) −f (b) < ε. By hypothesis, g (x) is
always in the domain of f, so all we need to do is to choose δ > 0 such
2. CONTINUITY OF NUMERICAL FUNCTIONS 31
that every x ∈ (c −δ, c +δ) ∩ E satisﬁes g (x) ∈ (b −δ
0
, b +δ
0
). The
continuity of g at c is exactly the condition we need to assert that such a δ
exists.
By applying these theorems in various combinations, we can recognize
the continuity of many functions deﬁned by elementary formulas. In par
ticular, if the values of a numerical function are determined by performing
the same ﬁxed, ﬁnite sequence of arithmetic operations at all points in the
domain, then the function must be continuous. We won’t try to write out
a formal proof of this general principle; the difﬁculty in proving it has
nothing whatever to do with the deﬁnition of continuity. Given any such
function, proving that it is continuous becomes a straightforward process
once we break it down to its simple parts. Rather, the difﬁculty in proving
the general principle is in characterizing those functions in such a way that
the argument given clearly applies to all of them.
When we’re trying to establish the continuity of a single function that
we don’t immediately recognize as continuous, we often study the rule
deﬁning the function and try to bound f (x) −f (a) in terms of x −a.
The general principle involved—which we often use with no conscious
effort—deserves to be stated as a theoremeven though its proof is extremely
simple.
Theorem 2.3: Suppose that f is a numerical function with domain
E, a ∈ E, and g is another numerical function such that
f (x) −f (a) ≤ g (x −a) for all x ∈ E.
If g is continuous at 0 with g (0) = 0, then f is continuous at a.
Proof : All we have to do is recognize that when ε > 0 is given,
the δ given by the continuity of g at 0 can serve as the δ needed for conti
nuity of f at a.
The deﬁnitions of simple functions often incorporate operations other
than addition, subtraction, multiplication, and division, such as deciding
such as whether x > a is true. That decision and others like it can
introduce discontinuities; every interval (a −δ, a +δ) contains both x for
which x > a is true and x for which it is false. Such decisions can present
practical difﬁculties when we attempt to evaluate the function at a value
determined only approximately, but they offer no logical problems and are
frequently used in mathematical theories.
According to what we said earlier, 1/x should be considered a contin
uous function. Its domain is the set of all nonzero real numbers and it’s
continuous at each point in its domain. But in common usage, we also say
32 CHAPTER II CONTINUOUS FUNCTIONS
that 1/x is discontinuous at x = 0, and it seems contradictory to talk about
the discontinuities of continuous functions. The reason for the apparent
paradox is that we haven’t yet given a precise deﬁnition of discontinuity.
In common usage, the notions of discontinuity at a point and continuity
at a point aren’t quite exact opposites. When we’re considering functions
that aren’t deﬁned on all of R, we sometimes consider points outside the
domain to be discontinuities.
The reader should already recognize that a numerical function f with
domain E is discontinuous at a point a ∈ E if for some ε > 0 every interval
of the form(a −δ, a +δ) will contain a point x ∈ Ewith f (x) −f (a) ≥
ε, no matter how small we choose δ > 0. Indeed, to prove a given function
f is discontinuous at a given point a in its domain, we usually must produce
such an ε. But we also consider f to be discontinuous where points are
missing from its domain. That is, f is discontinuous at a point b not in
its domain E if every interval (b −δ, b +δ) contains points of E. The
function f deﬁned by f (x) = 1/x has this sort of a discontinuity at 0.
While 0 isn’t in the domain of f, every interval (−δ, δ) contains points that
are. While we consider 1/
√
x to be discontinuous at 0, we don’t consider
it to be discontinuous at −1 because an interval (−1 −δ, −1 +δ) need not
contain any points in the domain of 1/
√
x.
According to our deﬁnitions, a function is either continuous or discon
tinuous at any given point in the domain. While a point outside the domain
can never be a point of continuity, it may or may not be a discontinuity;
that’s determined by the relationship of the point to the domain. A contin
uous function can indeed have discontinuities, but only at points outside its
domain.
EXERCISES
4. Use the deﬁnition of continuity to prove that the numerical function
f deﬁned by f (x) = 2x for all x ∈ Ris continuous.
5. Use the deﬁnition of continuity to prove that if f is continuous at a
point a and f (a) > 0, then there is an open interval (a −δ, a +δ)
containing no x with f (x) ≤ 0.
6. Let g be the numerical function whose value at each x ∈ R is the
decimal value of x, rounded to the nearest hundredth. Show that g
has discontinuities by producing an a ∈ R and an ε > 0 for which
no corresponding δ > 0 exists. How small does ε need to be in this
case?
7. Why is the function g in the previous exercise continuous on the
interval (−.005, .005)?
3. THE INTERMEDIATE VALUE THEOREM 33
8. Suppose that f is a numerical function with domain E, and that a
is a point of continuity of f. Prove that there is an open interval I
containing a for which the set of values {f (x) : x ∈ I ∩ E} has a
supremum and an inﬁmum.
9. Prove that if f is an increasing function on (a, b) and the set of values
{f (x) : a < x < b} is also an interval, then f must be continuous on
(a, b). (The conclusion remains true when (a, b) is replaced by an
arbitrary interval, but more cases must be considered to prove it.)
3 THE INTERMEDIATE VALUE THEOREM
Functions deﬁned and continuous on an interval are often thought of as
having graphs that form continuous paths. This evokes the image of a
graph that can be sketched without lifting the pencil used to draw it. That’s
thought to be the reason we use the word continuous to describe such
functions, even though that’s not the mathematical meaning of the word.
The theorembelow, universally known as the intermediate value theorem,
is closely related to this naive description of continuity.
Theorem 3.1: Let f be a numerical function that is continuous on
[a, b]. Then for any number r between f (a) and f (b), there is a real
number c in [a, b] with f (c) = r.
The graphical interpretation of this theorem is so striking that students
instinctively accept it as true. This interpretation says that when the hor
izontal line y = r separates the points (a, f (a)) and (b, f (b)), the graph
of f must cross the line at some point whose xcoordinate is between a
and b. Obviously the graph must also cross any vertical line that separates
the points, but we don’t need a theorem to tell us that. Since we know the
xcoordinate we can just evaluate f to locate the point on a given vertical
line. The intermediate value theorem deals with something that can be a
real problem: ﬁnding a value of x that satisﬁes the equation f (x) = r. We
should not dismiss the theorem as obviously true on geometrical grounds,
because if the graph had any gaps it could easily appear on both sides of
a line without actually crossing it anywhere, and the actual deﬁnition of
continuous function doesn’t explicitly say anything about the presence or
absence of gaps in the graph. Continuous functions aren’t necessarily as
nice as the smooth curves we use to visualize their graphs, so in our proof
we’ll have to be careful that we don’t assume something that isn’t actually
part of the hypotheses.
Before we prove the theorem, let’s consider an example and see how
we might deal with it. Let f be the function deﬁned by f (x) = x
5
+x. It’s
34 CHAPTER II CONTINUOUS FUNCTIONS
continuous on all of R but we’ll focus our attention on [0, 1]. Obviously
f (0) = 0 and f (1) = 2, so a simple application of the intermediate value
theorem says there must be an x ∈ [0, 1] with x
5
+x = 1. Algebra doesn’t
offer any method for solving this equation to ﬁnd x, but with the help of a
calculator we can get good results with a trialanderror procedure.
An examination of the formula deﬁning f shows us that increasing x
always increases f (x), so it’s easy to bracket the solution: x is too big if
f (x) > 1 and too small if f (x) < 1. We ﬁrst try the midpoint of [0, 1].
Since f (0.5) < 1 we narrow our search for x to [0.5, 1]. The midpoint of
this interval is 0.75, and since f (0.75) < 1, we narrow our search for x
to [0.75, 1]. At the next midpoint we ﬁnd f (0.875) > 1, so x must be in
[0.75, 0.875]. This process produces a nested sequence of closed intervals,
and the value of x we want must be the one number common to all of them.
Watching the trial values of f approach 1 provides strong evidence that x
really will have f (x) = 1.
What if we try to adapt this argument to an unspeciﬁed function?
The idea of ﬁnding c by producing a nested sequence of closed intervals
certainly looks appealing, but we’ll need some modiﬁcations. In particular,
an unspeciﬁed function might have either f (a) < r < f (b) or f (b) < r <
f (a) and still satisfy the hypotheses of the theorem, so we have to deal with
both possibilities. But a much greater problem is that we won’t actually
look at any numerical values of f, so we’ll need some other consideration
to explain why the c we ﬁnd really will satisfy f (c) = r. That’s where the
precise deﬁnition of continuity will come into play.
Proof : For the general case, we ﬁrst produce a nested sequence
of intervals
{[a
n
, b
n
]}
∞
n=1
,
starting with a
1
= a and b
1
= b. All of the different cases with r
between f (a
n
) and f (b
n
) can be summarized in a single inequality, one
that includes the cases f (a
n
) = r or f (b
n
) = r as well:
P
n
= [f (a
n
) −r] [f (b
n
) −r] ≤ 0. (2.1)
That inequality will also help us prove that the c we ﬁnd will have
f (c) = r. For n = 1 we have P
1
≤ 0 by hypothesis, and we pro
duce our sequence inductively in a way that makes sure P
n
≤ 0 for all
n. Let’s assume that [a
k
, b
k
] has been chosen with P
k
≤ 0, then call
m
k
the midpoint of [a
k
, b
k
]. We’ll choose [a
k+1
, b
k+1
] = [a
k
, m
k
] unless
[f (a
k
) −r] [f (m
k
) −r] > 0, and in that case we’ll choose [a
k+1
, b
k+1
] =
[m
k
, b
k
] instead. This procedure guarantees that P
k+1
≤ 0. Note that
EXERCISES 35
[f (a
k
) −r] [f (m
k
) −r] > 0 implies [f (m
k
) −r] [f (b
k
) −r] ≤ 0 since
their product is [f (m
k
) −r]
2
P
k
and we assumed that P
k
≤ 0.
Nowlet c be the one number common to all these intervals; it’s certainly
a point in our original interval [a, b]. To ﬁnish the proof we must show
that f (c) = r, and we do it by showing that f (c) = r requires f to have
a discontinuity at c. Our hypotheses include the continuity of f on all of
[a, b], so we can then conclude that f (c) = r.
For f (c) = r, every number close enough to [f (c) −r]
2
must be
positive, so our rules for approximate arithmetic say there is an ε > 0 such
that
(u −r) (v −r) > 0 for all u, v ∈ (f (c) −ε, f (c) +ε).
But no matter how small we pick δ > 0, we can ﬁnd n large enough that
both a
n
and b
n
are in (c −δ, c +δ) ∩ [a, b], and since P
n
≤ 0 we know
that one or both of the numbers f (a
n
), f (b
n
) must be outside the interval
(f (c) −ε, f (c) +ε). Thus if f (c) = r then f must have a discontinuity
at c, and that completes the proof.
We may note that if r is between f (a) and f (b) but doesn’t equal either
of them, then of course the c we ﬁnd with f (c) = r can’t be either a or b,
and therefore satisﬁes a < c < b. Consequently, there’s no reason to worry
about whether the hypothesis that r is between f (a) and f (b) includes the
case of equality; the versions of the theorem with either interpretation are
clearly equivalent.
We should note that the intermediate value theorem gives only a prop
erty that continuous functions have, not a deﬁnition of continuity. The
standard example of a discontinuous function that satisﬁes the conclusion
of the intermediate value theorem is
f (x) =
_
sin(1/x) , x > 0
0, x ≤ 0.
In every interval (0, δ), f takes on every value in [−1, 1], and that makes it
easy to see that intermediate values will always exist even though f is not
continuous at 0. An even more striking example is given in the book by
R. P. Boas [1]; he deﬁnes a function f such that for every open subinterval
(a, b) of [0, 1], the set of values {f (x) : a < x < b} is the entire interval
[0, 1].
EXERCISES
10. Suppose that f is a continuous numerical function on an interval I
and that V is the set of values {f (x) : x ∈ I}. Prove that V is an
36 CHAPTER II CONTINUOUS FUNCTIONS
interval. (It makes no difference whether the interval I is open or
closed, bounded or unbounded. Don’t try to determine the type of
intervals involved.)
11. Suppose that f : R →{0, 1} and that I is an interval on which f has
both values. Prove that f has a discontinuity in I.
12. Suppose that f is a continuous function on an interval I, a ∈ I, we’ve
been given an ε > 0, and we’re actually trying to calculate a positive
δ such that every x ∈ (a −δ, a +δ) ∩I satisﬁes f (x) −f (a) < ε.
Show that δ is suitable if (a −δ, a +δ) contains no x with f (x) =
f (a) ±ε.
13. For f (x) = 1/x and a > 0, ﬁnd the largest possible δ with the
property that every x ∈ (a −δ, a +δ) satisﬁes 1/x −1/a < ε by
solving
1/ (a +δ) = 1/a −ε and/or 1/ (a −δ) = 1/a +ε.
Finding δ is easy; explain how you know it will have the desired
properties.
14. If f and g are continuous numerical functions on [a, b] with f (a) <
g (a) but f (b) > g (b), why must there be an x ∈ (a, b) with f (x) =
g (x)?
4 MORE WAYS TO FORM CONTINUOUS FUNCTIONS
In addition to the algebraic formulas we use to deﬁne speciﬁc functions,
we sometimes consider functions for which the method used to determine
f (x) depends on whether x is in one or more specially designated sets.
While purely algebraic formulas generally produce only functions that are
continuous at all points in their domain, a formula that incorporates a
decision as to whether or not a value of the variable has a speciﬁc property
can easily lead to discontinuities. You may have noticed that such decision
rules have a way of appearing whenever anyone gives an example of a
function with a discontinuity at one or more points in its domain. However,
decision rules don’t automatically lead to discontinuities; let’s look at some
examples of what can happen.
The simplest continuous function involving a decision rule is the ab
solute value function,
x =
_
x, x ≥ 0
−x, x < 0.
4. MORE WAYS TO FORM CONTINUOUS FUNCTIONS 37
The absolute value function is continuous on all of R; indeed, for any ε > 0
we know every x ∈ (a −ε, a +ε) satisﬁes x −a < ε. Consequently,
when f is any numerical function we see that the function f deﬁned by
f (x) = f (x)
is continuous at every point where f is continuous. To satisfy the deﬁnition
of continuity, when ε > 0 is given we can use any δ for f that would work
for f.
Another way to look at the absolute value function is that evaluating it
amounts to choosing the larger of two numbers whose average is 0. We can
also use absolute values to select the greater or lesser of any two numbers a
and b. Since their average is halfway between them, both are at the distance
1
2
a −b from
1
2
(a +b), and so
max {a, b} =
1
2
(a +b) +
1
2
a −b,
min{a, b} =
1
2
(a +b) −
1
2
a −b.
That makes it easy to recognize that whenever f and g are continuous
functions on the same domain, selecting the larger of the two values f (x)
and g (x) at each point x deﬁnes another continuous function, and so
does selecting the smaller value at each point, because both
1
2
(f +g) and
1
2
f −g deﬁne continuous functions.
Our formula for the absolute value function is the most familiar example
of a useful technique: we take two or more functions with different domains
and patch them together to deﬁne a single function on the union of the two
domains. That is, we start with f : A → R and g : B → R, and deﬁne
h : A ∪ B → R so that h(a) = f (a) for a ∈ A and h(b) = g (b) for
b ∈ B. That works just ﬁne when A and B have no elements in common,
but it may not work when they do. For x ∈ A ∩ B, should h(x) be f (x)
or should it be g (x)? If the rule we use to decide depends on anything but
the value of x, it would seem to contradict the deﬁnition of function. But
in fact it may not; the deﬁnition requires that the value we assign to h(x)
depend only on the value of x; it says nothing about the process we use to
determine the value of h(x). In particular, if f (x) = g (x), then it doesn’t
really matter whether we say h(x) = f (x) or h(x) = g (x). That is, as
long as f (x) = g (x) for all x ∈ A ∩ B, we have no problem deﬁning
h : A∪ B →Rin terms of f and g. We call h the common extension of
f and g to the union of their domains.
What about the continuity of the common extension of two functions?
There are some simple implications here that we’ll state in a formal theo
38 CHAPTER II CONTINUOUS FUNCTIONS
rem. Its proof is quite easy, chieﬂy because it fails to address the difﬁcult
cases.
Theorem 4.1: For Aand B subsets of R, suppose that we have two
functions f : A →Rand g : B →Rwhose values agree on A∩B. Then
their common extension is continuous at each point where both f and g
are continuous, and any discontinuity of their common extension must be a
discontinuity for f or for g, if not for both.
Proof : Since every point in the domain of a function is either
a discontinuity or a point of continuity, the ﬁrst part of the theorem is a
logical consequence of the second so we’ll just prove the second part. To
ﬁx the notation, let’s call h the extension of f and g to A∪ B.
We begin by considering the possibilities at a point a / ∈ A∪B. Either
f has a discontinuity at a or there is a δ
1
> 0 such that (a −δ
1
, a +δ
1
)
contains no points of A, and either g has a discontinuity at a or there is a
δ
2
> 0 such that (a −δ
2
, a +δ
2
) contains no points of B. So when neither
f nor g has a discontinuity at a, we can take δ to be the smaller of δ
1
and δ
2
to get an interval (a −δ, a +δ) that doesn’t contain any points of A ∪ B.
This shows that h doesn’t have a discontinuity at a, and we’ve taken care
of the case a / ∈ A∪ B.
Next let’s suppose that a ∈ A and that h has a discontinuity at a.
There is nothing to prove if f has a discontinuity at a, so we might as well
assume that f is continuous at a. Since a is in the domain of h, there must
be an ε > 0 such that every interval (a −δ, a +δ) contains at least one
x ∈ A∪ B with
h(x) −h(a) ≥ ε.
But f is continuous at a, so when δ is small enough we know that every x
in A∩(a −δ, a +δ) satisﬁes f (x) −f (a) < ε. Since f and h agree on
A, any x ∈ (a −δ, a +δ) with h(x) −h(a) ≥ ε must be in B, not A.
Hence every open interval centered at a must contain x ∈ B with
g (x) −h(a) ≥ ε.
Since h(a) agrees with g (a) when a ∈ B, it follows that g has a disconti
nuity at a, whether or not a ∈ B.
The case a ∈ B can be argued similarly; we just interchange the roles
of f and g. That concludes the proof.
We routinely use this theorem to check the continuity of a class of
custombuilt functions commonly known as spline functions. We take
continuous functions deﬁned on adjacent intervals and patch themtogether.
4. MORE WAYS TO FORM CONTINUOUS FUNCTIONS 39
If the intervals are closed intervals and the values of the functions match at
the endpoints, then their common extension is necessarily continuous.
It’s entirely possible for the common extension of two functions to be
continuous at a point where one or the other has a discontinuity, as long
as the point isn’t in the domain of both. For example, if f (x) = x
2
/x
for x = 0 and g (x) = x for x ≥ 0, then their common extension is
continuous everywhere, even though f has a discontinuity at the origin.
Of course, one could argue that such a discontinuity shouldn’t even count.
Such discontinuities are called removable. Here’s a formal deﬁnition of
that concept.
Definition 4.1: A function f has a removable discontinuity at
the point a if f has a discontinuity there, but there is a function g that is
continuous at a and that agrees with f everywhere else in the domain of f.
In other words, a removable discontinuity can be transformed into a
point of continuity by either deﬁning or correcting the deﬁnition of the
function at the point. No points need to be deleted from the domain and
the values of the function need not be changed at any points except the one
where the removable discontinuity occurs. We call such a modiﬁcation
removing the discontinuity; it can be as simple as rewriting
x
2
−1
x−1
as x +1
to remove the discontinuity at x = 1. On the other hand, if f has a dis
continuity at a and there is an ε > 0 such that every interval (a −δ, a +δ)
contains a pair of points x, y = a with f (x) −f (y) ≥ 2ε, then the
discontinuity is certainly not removable. However f (a) is deﬁned, one or
both of the numbers f (x) −f (a) and f (y) −f (a) must be at least ε.
Basic calculus includes many problems that can be solved by recognizing
discontinuities as removable and then removing them. The next chapter
deals with the abstract theory behind removing discontinuities.
The reader should have noticed that the precise domain of the function
being considered has a great deal to do with answering questions of conti
nuity. On many occasions we have no interest in the values of a numerical
function at all points of its domain, but only at points in some subset. In
formal mathematics, it is then common to introduce a new function, called
the restriction of the original function to the set of interest. That is, when
A is a subset of the domain of f, the restriction of f to A keeps the same
values as f at each x ∈ A, but we think of A as its entire domain. For
example, if we deﬁne f on all of Rby the simple rule
f (x) =
_
0, x < 0
1, x ≥ 0,
40 CHAPTER II CONTINUOUS FUNCTIONS
then f has a discontinuity at the origin. However, the restriction of f to
[0, ∞) is continuous; the origin becomes a point of continuity when we
change the domain. In calculus we are often interested only in the values of
a function on a particular interval, but since we seldom specify the domain
of the functions we work with, we’ll adopt the following convention.
Definition 4.2: We say that the numerical function f is continuous
on the interval I if the domain of f includes I and the restriction of f to
I is a continuous function.
EXERCISES
15. Find a continuous numerical function on [0, 2] whose graph includes
the points (0, 2), (1, 0), and (2, 1) and follows straightline paths
between them.
16. Use the deﬁnition of continuity at a point to prove directly that the
common extension of f and g is continuous at each point where both
are continuous.
17. Write out a proof that if the common extension of f and g has a
discontinuity at a point where g is continuous, then f must have a
discontinuity there. (This was the last case considered in the theorem
proved in this section.)
18. The function f deﬁned by f (x) = x/ x has a discontinuity at the
origin. Either show why this discontinuity is not removable, or ﬁnd a
way to remove it.
19. Find the discontinuities of the function deﬁned by the formula
x
2
+x −6
x
3
−3x
2
+ 2x
.
Which are removable?
5 EXTREME VALUES
Finding the greatest or least value of a function is an inherently difﬁcult
procedure to perform analytically because it involves considering all the
values simultaneously, while analytic methods generally consider only one
value at a time. Looking at the values one at a time won’t even let you
determine whether the set of values has upper or lower bounds, and if the
set of values does have a supremum or an inﬁmum there is still the question
of determining whether they are actually values of the function. With many
functions, there are ways to get enough information about their graphs to
5. EXTREME VALUES 41
determine the extreme values, but in fact there is no general method for
doing so. Consequently, in this section we will often be able only to prove
that certain things exist without necessarily having a workable procedure
for ﬁnding them. That stands in stark contrast to our earlier work of proving
the existence of intermediate values by creating a procedure for producing
them.
Let’s consider the problemof ﬁnding the minimumvalue of a numerical
function f on a closed interval [a, b]; the problem of ﬁnding the maximum
is not very different. First assume that the minimum occurs at a point
c ∈ [a, b]. If we knew the value of c, it would be easy enough to produce
a nested sequence {[a
n
, b
n
]}
∞
n=1
of closed subintervals of [a, b] with c
the one number common to all of them. We could take [a
1
, b
1
] = [a, b]
and when [a
n
, b
n
] has been chosen with c ∈ [a
n
, b
n
], we then deﬁne
m
n
=
1
2
(a
n
+b
n
) and take
[a
n+1
, b
n+1
] =
_
[a
n
, m
n
], c ≤ m
n
[m
n
, b
n
], c > m
n
.
We wouldn’t really need to do any of this if we knew c, but the idea
suggests an interesting possibility: can such a sequence be used to prove
the existence of c?
It is indeed possible, but if we are going to use the sequence to prove
the existence of c then we must describe the sequence in a way that makes
no direct reference to either c or its existence. Instead of asking that
each interval [a
n
, b
n
] contain a single number c with the property that
f (c) ≤ f (x) for every x ∈ [a, b], we ask that for each x ∈ [a, b] there
be an element x
n
∈ [a
n
, b
n
] with f (x
n
) ≤ f (x). Such a sequence of
intervals must exist for any numerical function on any closed interval [a, b],
and when f is continuous on [a, b] its minimum value must occur at the one
point common to all of the intervals in the sequence. We’ll show how this
works as we prove the theorembelow. It seems both natural and convenient
to refer to the theorem as the extreme value theorem, but this name is not
in common use.
Theorem 5.1: Every numerical function that is continuous on a
closed interval has both a greatest and a least value on that interval.
Proof : We’ll only show why f has a minimum value; the maxi
mum value could always be found in terms of the minimum of −f. We use
induction to prove the existence of a sequence of intervals like we described
above. Getting the ﬁrst interval in the sequence is easy: [a
1
, b
1
] = [a, b].
Given x ∈ [a, b] we can choose x
1
= x ∈ [a
1
, b
1
] to get f (x
1
) ≤ f (x), so
[a
1
, b
1
] has the property we’re looking for. To prove the existence of the en
42 CHAPTER II CONTINUOUS FUNCTIONS
tire sequence, we assume that [a
n
, b
n
] is a closed subinterval of [a, b] with
length 2
1−n
(b −a) and that for each x ∈ [a, b] there is an x
n
∈ [a
n
, b
n
]
with f (x
n
) ≤ f (x). Then we showthat [a
n
, b
n
] must contain a subinterval
[a
n+1
, b
n+1
] with length 2
−n
(b −a) and having the property that for each
x ∈ [a, b], some x
n+1
∈ [a
n+1
, b
n+1
] satisﬁes f (x
n+1
) ≤ f (x).
As usual, we call m
n
the midpoint of [a
n
, b
n
]. The next interval
can be [a
n
, m
n
] unless some x
∗
∈ [a, b] satisﬁes f (x
∗
) < f (x) for all
x ∈ [a
n
, m
n
], and it can be [m
n
, b
n
] unless some x
∗∗
∈ [a, b] satisﬁes
f (x
∗∗
) < f (x) for all x ∈ [m
n
, b
n
]. If such x
∗
and x
∗∗
both existed, the
smaller of the values f (x
∗
) and f (x
∗∗
) would be strictly less than every
value of f on [a
n
, b
n
], violating our assumption about [a
n
, b
n
]. So we see
that a suitable [a
n+1
, b
n+1
] must exist.
Now we use continuity to prove that f must have its minimum value
on [a, b] at the one point c common to all the intervals in our nested se
quence. Obviously, f (c) is a value of f on [a, b]. We prove that it is
the minimum value by showing that no y < f (c) can be a value of f (x)
with x ∈ [a, b]. Given y < f (c), we choose ε > 0 with y < f (c) − ε.
By the deﬁnition of continuity at c, there must be a δ > 0 such that
every x ∈ (c −δ, c +δ) ∩ [a, b] satisﬁes f (x) −f (c) < ε and hence
f (x) > y. But when n is large enough to make b
n
−a
n
< δ we must have
[a
n
, b
n
] ⊂ (c −δ, c +δ). No x
n
∈ [a
n
, b
n
] could ever satisfy f (x
n
) ≤ y.
Thus y can’t be f (x) for any choice of x ∈ [a, b], and that proves the
theorem.
Now let’s turn our attention to another sort of extreme value problem.
If we assume that f is a continuous numerical function on [a, b], then for
ε > 0 and each c ∈ [a, b], we know that there must be a δ > 0 such that
every x ∈ (c −δ, c +δ) ∩[a, b] satisﬁes f (x) −f (c) < ε. We may also
note that the smaller the slope of the graph near (c, f (c)), the larger δ can
be. Conversely, where the slope is large we may need smaller values for
δ. But there is no bound on how steep the slope of the graph might get, so
it appears that for a given ε we might need to use arbitrarily small positive
values for δ. The theorem below shows that’s not the case when we restrict
our attention to a closed interval on which f is continuous; for each ε > 0,
a single value of δ can be used at all points in the interval.
Theorem 5.2: Let f be a numerical function that is continuous on
[a, b]. Then given any ε > 0, there is a δ > 0 such that f (x) −f (y) < ε
for every x and y in [a, b] with x −y < δ.
The conclusion of this theorem involves a condition that is so useful
that we give it a special name; we say that the function is uniformly
5. EXTREME VALUES 43
continuous. Explicitly, when f is a numerical function with domain E,
we say that f is uniformly continuous if for each ε > 0 there is a δ > 0
such that every x and y in E with x −y < δ satisfy f (x) −f (y) < ε.
Proof : We begin our proof by restating the conclusion: there is a
number δ > 0 such that every subinterval [y, z] of [a, b] with f (y) −f (z)
≥ ε has length at least δ. Let’s give a name to the subintervals we have to
watch out for: call
E = {[y, z] ⊂ [a, b] : f (y) −f (z) ≥ ε}.
Any δ will work if E = ∅, so we may as well assume that E = ∅. In that
case, we must prove there is a positive number δ that is less than or equal
to the length of every interval in E.
For E = ∅, we can deﬁne a numerical function ∆ on [a, b] by the
formula
∆(x) = inf {r > 0 : (x −r, x +r) has a subinterval in E}.
Clearly, 0 ≤ ∆(x) ≤ b − a for each x ∈ [a, b]. For any [y, z] ∈ E, we
also see that ∆(x) ≤ y − z at each x ∈ [y, z], so the δ we need can be
any number such that δ ≤ ∆(x) for all x ∈ [a, b]. So if we prove that ∆
is continuous on [a, b] and all its values are positive, we can appeal to the
extreme value theorem and call δ the minimum value of ∆.
In fact, ∆ is continuous on [a, b] whether or not f is continuous. To
prove it, we go back to the deﬁnition of ∆ and ﬁnd a way to compare its
values at two points x and x
. For any r > 0, we have
(x −r, x +r) ⊂
_
x
−
¸
¸
x −x
¸
¸
−r, x
+
¸
¸
x −x
¸
¸
+r
_
, (2.2)
so if the interval on the lefthand side of (2.2) has a subinterval in E then
so does the one on the right. Consequently,
∆
_
x
_
≤
¸
¸
x −x
¸
¸
+ ∆(x),
and since the roles of x and x
can be switched, we ﬁnd
¸
¸
∆(x) −∆
_
x
_¸
¸
≤
¸
¸
x
−x
¸
¸
.
That actually shows that ∆is uniformly continuous, but that isn’t important
here. All we wanted is simple continuity.
When f is continuous on [a, b], then ∆(c) is positive at each c. In fact,
∆(c) ≥ r whenever r is small enough that every x ∈ (c −r, c +r) ∩[a, b]
satisﬁes f (x) −f (c) < ε/2. After all,
f (y) −f (z) ≤ f (y) −f (c) +f (z) −f (c),
44 CHAPTER II CONTINUOUS FUNCTIONS
so one or both of the endpoints of any interval in E must fall outside the
interval (c −r, c +r). Hence the minimum value of ∆ is positive, and that
completes the proof.
Here’s an important application of the last theorem. Suppose that we’re
interested in plotting the graph of a numerical function f that is continuous
on [a, b], and we want the ycoordinates of our graph to be accurate to
within ε at all points of that interval. If we ﬁnd the number δ given by the
theorem, we can then pick an integer n with nδ > b −a. Then we call
x
k
= a +
k
n
(b −a) for k = 0, 1, . . . , n.
For x ∈ [x
k−1
, x
k
], we’ll have x
k−1
−x < δ and x
k
−x < δ. So
both f (x
k−1
) and f (x
k
) will be in (f (x) −ε, f (x) +ε). Every number
between f (x
k−1
) and f (x
k
) must also be in (f (x) −ε, f (x) +ε). So
the brokenline graph we obtain by successively connecting the points
{(x
k
, f (x
k
))}
n
k=0
will have the desired accuracy.
EXERCISES
20. Prove that if f is a continuous numerical function on a closed interval
I, then the set of values {f (x) : x ∈ I} is itself a closed interval.
21. Modify the argument given in the proof of the extreme value theorem
to prove the existence of a greatest value.
22. In the case of
√
x on [0, 1] with ε given in (0, 1), show that ε
2
is the
largest value of δ satisfying the conclusion of Theorem 5.2.
23. If a numerical function f is increasing and continuous on [a, b], the
intermediate value theorem can be used to ﬁnd the δ in Theorem 5.2.
Choose an integer n larger than 2 [f (b) −f (a)] /ε, and explain why
for k = 0, 1, 2, . . . , n it is possible to ﬁnd x
k
in [a, b] such that
f (x
k
) = f (a) +
k
n
[f (b) −f (a)].
Explain why the least of the numbers {x
k
−x
k−1
: 1 ≤ k ≤ n} is a
suitable δ.
III
LIMITS
W
hat we can actually do in mathematics is an incredibly small
part of what we would like to do. Operations we can per
form on numbers don’t really go much beyond the addition,
subtraction, and multiplication of whole numbers; we are limited in what
we do by a fundamental restriction. Namely, we must be able to ﬁnish the
job. If we can’t ﬁnish it, then it is not something we can actually do, even
if we’re capable of performing each step the job requires. This is the sort
of limitation we run into in trying to calculate
√
2, for example. But the
mere fact that we humans can’t complete a process doesn’t stop us from
studying it, theorizing about it, and trying to ﬁnd what would result if it
were somehow completed. By studying the theoretical result of processes
we can’t complete, we’ve made many important discoveries. Sometimes
the obstacle is more imagined than real, and while it blocks an approach it
may leave the goal itself accessible.
Calculus provides us with a systematic way to transcend the limitations
on the calculations we can complete, and that makes it one of the greatest
45
46 CHAPTER III LIMITS
intellectual achievements of our civilization. The one concept that makes
all this possible is the notion of a limit, a word we use to describe some
operations that appear to involve a step just beyond what we can complete.
That is, while we may never be able to complete all the steps required to
calculate limits, we can approximate them arbitrarily closely by quantities
that we do have methods for computing. We discover that it is quite
possible to work with limits, and it is often even a simple matter to ﬁnd
exact relationships between quantities deﬁned in terms of them.
Many different sorts of limits are used in calculus. They amount to
simple variations on a basic theme, and most of them come quite readily
once a single basic notion of limit has been mastered. For that reason,
we’ll begin our study of limits with the one that is easiest to understand:
the limit of a sequence of numbers.
1 SEQUENCES AND LIMITS
Let {x
n
}
∞
n=1
be a sequence of real numbers. That is, for each natural
number n there is a real number x
n
deﬁned. We do not concern ourselves
with how the process deﬁning the sequence works. But we are concerned
with properties of the sequence as a whole, not just with individual terms.
In principle there is no obstacle to examining the beginning terms of the
sequence one at a time, no matter how many we have to look at; it’s dealing
with the rest of the terms that’s a challenge. We call a set {x
n
: n ≥ m} a
tail of the sequence, and this is the part that holds our interest for now. If
we can ﬁnd some way to deal with a tail all at once, then studying properties
of the entire sequence presents no insurmountable difﬁculty.
Convergent sequences have such a property; they have tails in which
all the terms are nearly equal. Let’s explore this notion further before
we attempt to formulate a precise deﬁnition. Assuming that {x
n
}
∞
n=1
is a
bounded sequence of real numbers, we can use the tails to deﬁne a sequence
of intervals {[a
m
, b
m
]}
∞
m=1
by setting
a
m
= inf {x
n
: n ≥ m} and b
m
= sup {x
n
: n ≥ m}.
Recall that we don’t call [a
m
, b
m
] an interval unless a
m
≤ b
m
, and we
know that’s true because a
m
≤ x
m
≤ b
m
for each m. The sequence of
intervals we’ve deﬁned is actually a nested sequence of closed intervals.
They’re nested because a
m
and b
m
are bounds for {x
n
: n ≥ m+ 1} as
well as for {x
n
: n ≥ m}, so a
m
≤ a
m+1
and b
m+1
≤ b
m
. Each [a
m
, b
m
]
is the smallest closed interval that contains the tail {x
n
: n ≥ m}. The
diameter of the tail is the length of the interval, so b
m
− a
m
tells us how
much variation there is among its terms in the tail. We’ll want the deﬁnition
1. SEQUENCES AND LIMITS 47
of convergent sequence to imply that b
m
−a
m
→ 0. When that’s the case,
our axioms guarantee that there is exactly one number L common to all
the intervals {[a
m
, b
m
]}
∞
m=1
. Every term in the tail {x
n
: n ≥ m} must be
within b
m
−a
m
of L, so in that case we can indeed think of every term in
the tail {x
n
: n ≥ m} as nearly the same as Lwhenever mis large enough.
Nowlet’s give a formal deﬁnition that pins down these ideas. It’s better
if our deﬁnition omits any mention of the nested sequence {[a
m
, b
m
]}
∞
m=1
because we may not have an effective method for calculating the supremum
and inﬁmum of any of the tails, let alone all of them.
Definition 1.1: A sequence {x
n
}
∞
n=1
of real numbers is a con
vergent sequence if there is a number L with the property that for each
ε > 0 there is an integer M such that every x
n
with n ≥ M satisﬁes
x
n
−L < ε. The number L is called the limit of the sequence {x
n
}
∞
n=1
,
and it is common to say that the sequence converges to its limit. We indicate
this special relationship between the sequence and its limit symbolically as
lim
n→∞
x
n
= L.
When we’re primarily interested in displaying the value of the limit but
not the nature of the limiting process involved, we may abbreviate it as
x
n
→ L.
In other words, a sequence converges to a limit if every open interval
containing the limit also contains a tail of the sequence. A sequence can’t
possibly converge to more than one limit because we can certainly ﬁnd
disjoint open intervals around any two distinct numbers, and if one of those
intervals contains a tail then the other one can’t.
Now let’s see that the deﬁnition we gave really does correspond to the
ideas we discussed before giving it. We began our discussion by assuming
that {x
n
}
∞
n=1
was a bounded sequence, but that doesn’t seem to be part of
the deﬁnition given. It is, however, a consequence, which is why it wasn’t
stated explicitly. For example, if we choose ε = 1 and then ﬁnd M such
that every x
n
with n ≥ M satisﬁes x
n
−L < 1, then x
n
 ≤ L + 1 for
all n ≥ M, and we get a bound for the entire sequence:
x
n
 ≤ max {x
1
, x
2
, . . . , x
M−1
, L + 1} for all n ∈ N.
So every convergent sequence is necessarily bounded. Returning to the
case of arbitrary ε, if every x
n
with n ≥ M satisﬁes x
n
−L < ε, then
L − ε and L + ε are lower and upper bounds for the tail {x
n
: n ≥ M}.
So we’ll have
L −ε ≤ inf {x
n
: n ≥ M} = a
M
≤ b
M
= sup{x
n
: n ≥ M} ≤ L +ε.
48 CHAPTER III LIMITS
Thus b
M
−a
M
< 2ε, and since ε is arbitrary our nested sequence of closed
intervals will indeed satisfy b
m
−a
m
→ 0.
We can combine convergent sequences in a variety of ways to produce
newones. For example, if {x
n
}
∞
n=1
and {y
n
}
∞
n=1
are convergent sequences
with x
n
→ L and y
n
→ L
, then x
n
+ y
n
→ L + L
and x
n
y
n
→
LL
; those are consequences of the rules for approximate arithmetic. For
example, if we’re given ε > 0 and we want x
n
y
n
−LL
 < ε, we know
there is a δ > 0 such that it’s true whenever x
n
−L < δ and y
n
−L
 <
δ. Invoking the assumed convergence of the two original sequences, there
must be integers M and M
such that every x
n
with n ≥ M satisﬁes
x
n
−L < δ and every y
n
with n ≥ M
satisﬁes y
n
−L
 < δ. Both
conditions are satisﬁed for n greater than or equal to the larger of M and
M
, which proves x
n
y
n
→ LL
.
In addition to combining convergent sequences arithmetically to pro
duce new convergent sequences, we may also use continuous functions for
the same purpose. Let’s look at how it works. Suppose that x
n
→ L and
that f is a numerical function whose domain includes L as well as every
termin the sequence. Then {f (x
n
)}
∞
n=1
is a newsequence of real numbers,
and f (x
n
) → f (L) if f is continuous at L. Again, this is quite easy to
prove. When we’re given ε > 0, we know there is a δ > 0 such that every
x in the domain of f with x −L < δ must satisfy f (x) −f (L) < ε.
For x
n
→ L, there is an integer N such that x
n
∈ (L −δ, L +δ) for all
n ≥ N, and we’ve explicitly assumed that each x
n
is in the domain of f.
Since we can work with convergent sequences so easily, most of the
sequences we use will be convergent ones. But the ones we use are created
with some care. A sequence constructed with less care may not converge,
and then we call it divergent. There are many ways that a sequence of real
numbers can fail to converge. There are certainly unbounded sequences,
and since every convergent sequence is bounded we know that no un
bounded sequence of real numbers can possibly converge. But bounded
sequences of real numbers can also diverge; their terms can oscillate so that
the diameters of the tails all stay above some positive value. The sequence
deﬁned by x
n
= (−1)
n
is an especially simple example of a divergent
bounded sequence. More generally, any time we can ﬁnd an open interval
(a, b) such that every tail of the sequence includes both terms below a and
terms above b, the sequence can’t converge; for ε <
1
2
(b −a) the interval
(L −ε, L +ε) can never contain a tail of the sequence, no matter how we
choose L.
EXERCISES 49
EXERCISES
1. Let {x
n
}
∞
n=1
be a convergent sequence, and suppose that x
n
> 0 for
every n. Prove that lim
n→∞
x
n
≥ 0.
2. Let {x
n
}
∞
n=1
and {y
n
}
∞
n=1
be convergent sequences, with x
n
→ a
and y
n
→ b. What must we assume to prove that {x
n
/y
n
}
∞
n=1
is a
sequence that converges to a/b?
3. Let {x
n
}
∞
n=1
be a convergent sequence, and let I be an open interval.
Prove that if the sequence has inﬁnitely many terms that are not in I,
then lim
n→∞
x
n
/ ∈ I. Is there a similar result for closed intervals?
2 LIMITS AND REMOVING DISCONTINUITIES
Our condition for a discontinuity a of a numerical function f to be remov
able can be hard to check. We said that the discontinuity is removable
if there is a function g whose domain includes both a and the domain of
f, that is continuous at a, and that agrees with f everywhere except at a.
That’s ﬁne when we can spot a way to deﬁne g by simply rewriting the
formula for f, but how can we tell whether such a function g exists when
there’s no apparent formula for it? We really ought to have a condition
stated in terms of f, not some unknown function g. And we ought to
have a convenient way to refer to the value we assign to f (a) to remove
the discontinuity. That’s the idea behind the following deﬁnition. Since
our notion of continuity at a point depends on the point, the values of the
function, and the domain of the function, our deﬁnition also depends on all
these things.
Definition 2.1: Let f be a numerical function whose domain
includes the set E, and let a be a point such that every interval (a −δ, a +δ)
with δ > 0 contains points of E different from a. We say that the number
L is the limit of f(x) as x approaches a through E if for each ε > 0
there is a δ > 0 such that every x ∈ (a −δ, a +δ) ∩ E, with the possible
exception of a itself, satisﬁes f (x) −L < ε. We indicate this condition
symbolically by
lim
x→
E
a
f (x) = L.
The deﬁnition given in most calculus courses doesn’t involve the set
E, but there is a requirement that the domain of f include every x = a in
some open interval about a. However, the idea of restricting x to a smaller
subset of values appears in the guise of onesided limits, and that simple
model helps us understand the deﬁnition given here. We’ll return to the
notion of onesided limits at the end of this section.
50 CHAPTER III LIMITS
It’s important to recognize that our deﬁnition does specify the number
L we call the limit. That is, if we have a point a, a set E, a function
f whose domain includes E, and numbers L and L
that both satisfy the
deﬁnition of the limit of f (x) as x approaches a through E, then we must
have L = L
. In the previous section we used the fact that every open
interval about the limit of a sequence contains a tail of the sequence to
prove that no sequence could have more than one limit, and we would like
to do something similar here. In this case, sets of the form
{f (x) : x ∈ (a −δ, a +δ) ∩ E, x = a}
play the same role as the tail of a sequence. Of course, the tail of an
inﬁnite sequence will always have at least one point in it because there will
always be n ≥ m. Similarly, every set of the form given above will have
at least one point because of our assumption that every interval of the form
(a −δ, a +δ) contains points of E other than a.
It’s clear that if L is the limit of f (x) as x approaches a through E,
and we deﬁne g : E∪{a} →Rby g (x) = f (x) for x = a and g (a) = L,
then g is continuous at a and agrees with f at all other points of E ∪ {a}.
Consequently, when E is the entire domain of f and lim
x→
E
a
f (x) exists,
then either f is already continuous at a and g is the same as f or f has a
removable discontinuity at a and L is the value we use for f (a) to remove
it. Note that our deﬁnition makes no assumptions about whether a is a
point in the domain of f, and it makes no direct mention of f (a). That’s
as it should be; when f has a removable discontinuity at a, we think of
the value f (a) as either missing or incorrectly deﬁned. However, when
we know that f is continuous at a we can simply evaluate f (a) to ﬁnd the
limit.
The bad news about our deﬁnition is that it doesn’t tell us how to ﬁnd
L or even how to determine whether such a number can exist, and doing so
can be very hard. While we can determine limits in many speciﬁc cases, in
general the best we can do is to ﬁnd alternative descriptions of the limit and
the circumstances under which it exists. For example, when L is known
to be the limit of f (x) as x approaches a through E, then for sufﬁciently
small δ > 0 the set of values
{f (x) : x ∈ E with 0 < x −a < δ}
is a bounded, nonempty set of real numbers. For convenience, let’s call c
δ
its inﬁmum and d
δ
its supremum. Then the deﬁnition of limit guarantees
us that for any ε > 0, there is a δ > 0 such that c
δ
≥ L−ε and d
δ
≤ L+ε.
Conversely, the numbers c
δ
and d
δ
can be deﬁned without knowing that
2. LIMITS AND REMOVING DISCONTINUITIES 51
the limit exists. It’s not hard to show that the limit does exist if and only
if there is exactly one number common to all the intervals [c
δ
, d
δ
]; that one
number is the value of the limit. But that doesn’t really tell us how to ﬁnd
L because we may not know exactly how to locate any of the intervals
[c
δ
, d
δ
], let alone all of them.
We can ﬁnd statements of other conditions under which L must exist,
but unless we assume additional hypotheses we won’t have a systematic
way to determine whether they are satisﬁed in general. Nonetheless, ver
ifying such conditions for speciﬁc functions is a very real possibility. If
we’re clever enough, we can ﬁnd a way to complete the task for just about
every function that we can write down explicitly. But the methods we use
invariably make use of special properties of the function at hand, which we
must ﬁrst recognize and then ﬁnd a way to exploit. Recognizing an appro
priate technique may be quite difﬁcult, even if we understand the theory of
limits perfectly. We’re probably more likely to ﬁnd a limit by recognizing
a familiar way to remove a discontinuity than to remove a discontinuity by
analyzing a limit.
Something that is easy to do with limits and functions in general
is to recognize how the limits combine when we combine two or more
functions. The deﬁnition of limit is designed to simplify this task and
works beautifully with the rules for approximate arithmetic; most cases
reduce to a single argument that gets used over and over again. If, for
example, lim
x→
E
a
f (x) = L
1
and lim
x→
E
a
g (x) = L
2
, then it’s easy to
see that lim
x→
E
a
f (x) g (x) = L
1
L
2
. We can make f (x) g (x) −L
1
L
2

as small as necessary by simply making sure that f (x) is close enough to
L
1
and that g (x) is close enough to L
2
. Both of these conditions can be
achieved by taking x = a in E ∩(x −δ, x +δ) with an appropriate δ > 0.
Once again, the key is to recognize that if it’s possible to choose positive
δ’s to satisfy each of several conditions, then choosing the smallest of them
allows us to satisfy all the conditions simultaneously.
Analyzing the limit is a little more complicated when a function of a
function is involved. In considering lim
x→
E
a
f (x), nothing is assumed
about f (a), which allows for some nasty surprises. To rule out such
surprises, we need additional hypotheses beyond those we might expect,
and there are several useful possibilities. The theorem below is a good
example.
Theorem 2.1: Let A and B be sets of real numbers, with f :
A → B and g : B → R. Suppose that lim
x→
A
a
f (x) = b and that
lim
x→
B
b
g (x) = L. Then lim
x→
A
a
g (f (x)) = L, provided either that b
52 CHAPTER III LIMITS
is a point of continuity for g or that there is a δ
0
> 0 such that b is not one
of the values of f on the set {x ∈ A : 0 < x −a < δ
0
}.
Proof : We need to show that whenever ε > 0 is given, there is a
δ > 0 such that every x ∈ (a −δ, a +δ) ∩A except possibly a itself must
satisfy g (f (x)) −L < ε. Since we’ve assumed that L = lim
x→
B
b
(x),
we knowthat there is an η > 0 such that every y ∈ (b −η, b +η)∩Bexcept
possibly y = b satisﬁes g (y) −L < ε. While we can pick δ > 0 such
that every x ∈ Awith 0 < x −a < δ satisﬁes f (x) ∈ (b −η, b +η)∩B,
that may not be good enough. It is good enough when g is continuous at
b because then g (f (x)) = L when f (x) = b. Without continuity at b,
we need to be able to choose δ > 0 in such a way that no x ∈ A with
0 < x −a < δ satisﬁes f (x) = b.
Discovering and proving theorems about limits of combinations of
functions is basically a simple matter; all the difﬁculties in choosing a δ
for a given ε are buried in the hypotheses assumed. Using the theorems
is often so natural that as we break a complex function down into simpler
components, we may not even be aware of the theorems involved. The
skill comes in knowing when not to use the theorems; it’s important to
recognize when an expression with several parts must be dealt with as a
unit.
In elementary calculus, we don’t ordinarily deal with functions deﬁned
on arbitrary sets, and we’re generally interested in the limit of a function
at a point that is either an interior point or an endpoint of an interval in
the domain of the function. Consequently, lim
x→
E
a
f (x) is not ordinarily
used. When E includes all x = a in an open interval centered at a, the limit
in question is called simply the limit of f (x) as x approaches a, indicated
symbolically as lim
x→a
f (x) = L. We sometimes write this informally
as f (x) → L as x → a. When E is an interval (a, b), the limit is called
the righthand limit of f (x) as x approaches a, indicated symbolically as
lim
x→a+
f (x). When E is (c, a), the limit is called the lefthand limit,
indicated as lim
x→a−
f (x). But all of these are simply special cases of a
single concept, so a single theorem can deal with all three simultaneously,
as we’ve done.
In fact, we sometimes investigate lim
x→a
f (x) by considering the
lefthand and righthand limits separately. This is especially useful when
different computational formulas for f (x) are involved, one used for x < a
and the other for x > a. Of course, lim
x→a
f (x) need not exist, but when
it does it’s a single number and has to be equal to both lim
x→a+
f (x) and
lim
x→a−
f (x). While neither lim
x→a+
f (x) nor lim
x→a−
f (x) has to
exist in general, when both exist but have different values we say that f
EXERCISES 53
has a simple jump discontinuity at a. For example, x/ x has a simple
jump discontinuity at the origin. Such discontinuities are easy to deal with
as long as we stay on the same side of the jump since on either side they
act like removable discontinuities.
EXERCISES
4. Let f (x) be a nonnegative function, and suppose that a is a point
for which lim
x→a
f (x) exists. Prove that lim
x→a
f (x) ≥ 0. If we
make the stronger assumption f (x) > 0 for all x, can we conclude
that lim
x→a
f (x) > 0?
5. Prove that when lim
x→
E
a
f (x) = L
1
and lim
x→
E
a
g (x) = L
2
, we
must have lim
x→
E
a
[f (x) +g (x)] = L
1
+L
2
.
6. Prove that if p (x) is any polynomial, a is any real number, the function
f deﬁned by [p (x) −p (a)] / (x −a) has a removable discontinuity
at x = a. You’ll need to make use of an important principle from
algebra.
3 LIMITS INVOLVING ∞
As we study the graphs of given functions, we often check for the presence
of horizontal asymptotes. There’s an analytic version of this feature, too;
the notion of the limit of a function at ∞or −∞. When there are horizontal
asymptotes, the function values exhibit little change after we’ve moved far
to the right or left. Just as we can use lim
n→∞
x
n
to approximate all the
terms of the sequence {x
n
}
∞
n=1
with n > N, we can use lim
x→∞
f (x)
and lim
x→−∞
f (x) to approximate the values of f for all x outside a
closed interval [a, b]. That goal and our earlier work with limits pretty well
tell us how these limits need to be deﬁned, so we’ll give their deﬁnition
immediately.
Definition 3.1: We say that lim
x→∞
f (x) = L provided that for
each ε > 0, there is a b ∈ Rsuch that every x ∈ (b, ∞) is in the domain of f
and satisﬁes f (x) −L < ε. Similarly, we say that lim
x→−∞
f (x) = L
provided that for each ε > 0, there is a real number a such that every
x ∈ (−∞, a) is in the domain of f and satisﬁes f (x) −L < ε.
Of course, these limits don’t necessarily exist, but when they’re known
to exist we can use that information in many ways. For example, if two or
more functions are known to have limits at ∞or at −∞, we canimmediately
recognize the corresponding limits of many algebraic combinations of
them. The proofs of such relationships are not signiﬁcantly different from
54 CHAPTER III LIMITS
arguments given before; we just choose b large enough or a small enough
to satisfy several conditions simultaneously.
To evaluate expressions of the form lim
x→∞
[p (x) /q (x)] with p and
q speciﬁc polynomials, many students look at the leading terms of the
polynomials and ignore all the rest. Let’s look at what’s involved, and see
when that simpliﬁed approach can be successful.
For p (x) a polynomial of degree n, it is possible to write
p (x) = a
n
x
n
+a
n−1
x
n−1
+· · · +a
1
x +a
0
,
where the coefﬁcients {a
k
}
n
k=0
are constants with a
n
= 0. We can always
rewrite the formula for p (x) as
p (x) = x
n
_
a
n
+a
n−1
x
−1
+· · · +a
1
x
1−n
+a
0
x
−n
¸
= x
n
P (1/x),
where P is the polynomial deﬁned by
P (y) = a
n
+a
n−1
y +· · · +a
1
y
n−1
+a
0
y
n
.
We can’t say the degree of P is n unless a
0
= 0, but we’ve assumed that
P (0) = a
n
= 0.
Since all polynomials are continuous, P is continuous at 0. So when ε > 0
is given, there must be a δ > 0 such that
¸
¸
¸
¸
P (y)
P (0)
−1
¸
¸
¸
¸
< ε for all y ∈ (−δ, δ).
Consequently,
¸
¸
¸
¸
p (x)
a
n
x
n
−1
¸
¸
¸
¸
=
¸
¸
¸
¸
P (1/x)
a
n
−1
¸
¸
¸
¸
< ε for all x ∈ (1/δ, ∞),
which is the basis for approximating p (x) by a
n
x
n
as x → ∞. Their ratio
approaches 1, although their difference may be quite large.
The following theoremis a different sort of result that uses the existence
of limits at ±∞as a hypothesis and shows howwe use those limits to reduce
the study of a function on (−∞, ∞) to looking at what it does on a closed
interval [a, b].
Theorem 3.1: Let f be a continuous numerical function on R, with
lim
x→∞
f (x) = R and lim
x→−∞
f (x) = L. Then f is a bounded func
3. LIMITS INVOLVING ∞ 55
tion, with sup{f (x) : −∞ < x < ∞} and inf {f (x) : −∞ < x < ∞}
either R, L, or values of f.
Proof : Using ε = 1 in the deﬁnition of limit, we can choose a
and b such that every x in (−∞, a) satisﬁes f (x) −L < 1, and every x in
(b, ∞) satisﬁes f (x) −R < 1. Moreover, we can assume that a < b.
Since f is continuous on [a, b], we know there are numbers x
0
and x
1
in
that closed interval such that
f (x
0
) ≤ f (x) ≤ f (x
1
) for all x ∈ [a, b].
We also know that
L −1 < f (x) < L + 1 for all x ∈ (−∞, a)
as well as
R −1 < f (x) < R + 1 for all x ∈ (b, ∞).
Consequently, the least of the three numbers f (x
0
), L − 1, and R − 1 is
a lower bound for the values of f, and the greatest of the three numbers
f (x
1
), L + 1, and R + 1 is an upper bound.
Having proved that f is a bounded function, we can deﬁne
A = inf {f (x) : x ∈ R} and B = sup{f (x) : x ∈ R},
and then f (x) ∈ [A, B] for all x. The deﬁnition of limit won’t allow either
L or R to be below A or above B, so both must also be in [A, B]. Let’s
suppose that neither L nor R is B. In that case, we can ﬁnd ε > 0 such
that both L +ε < B and R+ε < B, and then we can choose a
such that
f (x) < L +ε < B for every x ∈
_
−∞, a
_
as well as b
such that
f (x) < R +ε < B for every x ∈
_
b
, ∞
_
.
Since f is continuous on [a
, b
], it has a greatest value on that interval, and
that value must be B. Similarly, if neither L nor R equals A, then we can
produce another interval [a
, b
] on which the least value of f is A.
We conclude this section by noting that sometimes limits are said to be
∞or −∞; these are referred to as inﬁnite limits. Some caution is needed
to deal with inﬁnite limits correctly. It is best to think of an inﬁnite limit as
only a convenient description of a property that a function without a limit
56 CHAPTER III LIMITS
may have. We can’t use our rules for approximate arithmetic to work with
algebraic combinations of functions with inﬁnite limits since the symbols
∞ and −∞ don’t represent real numbers and are not subject to the usual
laws of arithmetic. We can, however, use lim
x→∞
f (x) or lim
x→−∞
f (x)
to evaluate the limit of f (g (x)) when g has an inﬁnite limit; the deﬁnitions
of inﬁnite limits are designed to make this possible. Just as we say that
f (x) → L if for every ε > 0 there are appropriate intervals in which every
x satisﬁes f (x) ∈ (L −ε, L +ε), we say f (x) → ∞if for every b there
are appropriate intervals in which every x satisﬁes f (x) ∈ (b, ∞). We
make a similar modiﬁcation to deﬁne f (x) → −∞. Of course, the exact
deﬁnition depends on the sort of limit involved: twosided limits, onesided
limits, limits at −∞, or limits at ∞.
EXERCISES
7. For f (x) = xx /
_
x
2
+ 1
_
, ﬁnd the supremum and inﬁmum of the
set of values of f. Are they values of f?
8. Suppose that f is a numerical function deﬁned on (0, ∞). While the
limit of the sequence {f (n)}
∞
n=1
and lim
x→∞
f (x) often refer to
the same number, they are actually different concepts. What is the
relationship between them?
9. Prove that if lim
x→∞
f (x) = L
1
and lim
x→∞
g (x) = L
2
, then
lim
x→∞
[f (x) +g (x)] = L
1
+L
2
.
10. Prove that if lim
x→∞
f (x) = L, then lim
x→0+
f (1/x) = L as well.
11. Formulate a precise deﬁnition of lim
x→∞
f (x) = ∞.
12. Prove that if lim
x→∞
f (x) = L and lim
x→∞
g (x) = ∞, then
lim
x→∞
[f (x) +g (x)] = ∞.
13. Prove that if lim
x→∞
f (x) = ∞, then lim
x→∞
1/f (x) = 0.
IV
THE DERIVATIVE
S
ince calculus developed largely to meet the needs of science, there
are many parallels between scientiﬁc methodology and calculus
concepts. For the most part, science proceeds by studying change,
and the study of changes is the basis for differential calculus.
1 DIFFERENTIABILITY
The ﬁrst step in any serious study of change is to ﬁnd a way to measure
it. That’s especially easy to do with numerical functions. The quantity
f (x) − f (a) represents a change in the value of the function f, and
comparing this change to x −a gives us a convenient scale for measuring
it. Many elementary functions exhibit a common pattern:
f (x) −f (a) = (x −a) g (x), with g continuous at a. (4.1)
For example, if f is deﬁned by any polynomial, then f (x) − f (a) is a
polynomial that vanishes when x = a, and so by the factor theorem we
57
58 CHAPTER IV THE DERIVATIVE
know that (x −a) is a factor of f (x) −f (a). The complementary factor
g (x) in (4.1) is also a polynomial, and we know that polynomials deﬁne
continuous functions on the entire real line. When f isn’t a polynomial,
representations of the form (4.1) may still be possible, but the continuity
of g at a is harder to establish.
It is the continuity of g at a that makes equation (4.1) useful because
it lets us approximate g (x) by g (a) when x is near to a. To use this
approximation more effectively, we rewrite equation (4.1) as
f (x) = f (a) + (x −a) g (a) + (x −a) [g (x) −g (a)].
The sumf (a)+(x −a) g (a) forms the linearization of f (x) near x = a,
and the term (x −a) [g (x) −g (a)] is thought of as a correction to the
linearization. For g continuous at a, the correction term is small in
comparison to x −a when x is near a. More precisely, for any given
ε > 0 there is an interval (a −δ, a +δ) on which the correction is never
larger than ε x −a. The linearization of f is quite easy to understand and
to work with, so when we can ignore the effect of the correction term we
can eliminate all the difﬁculties that complicated functions can introduce.
This is the basic idea behind differentiability, the concept deﬁned below.
Definition 1.1: For f a numerical function, we say the function f
is differentiable at the point a if the domain of f includes an open interval
I about a, and there is a function g : I → R that is continuous at a and
satisﬁes
f (x) = f (a) + (x −a) g (x) for all x ∈ I.
When f is differentiable at a, the function g in our deﬁnition must be
deﬁned at each x = a in I by the formula
g (x) =
f (x) −f (a)
x −a
. (4.2)
We may note that whether or not f is differentiable at a, the function g
deﬁned by (4.2) is continuous at all the other points of continuity of f. Of
course, the quotient in (4.2) has a discontinuity at a since we’re assuming it
is deﬁned at all the other points in I. Then f is differentiable at a precisely
when we can remove the discontinuity at a. In that case,
g (a) = lim
x→a
f (x) −f (a)
x −a
. (4.3)
So another way to deﬁne differentiability is to say that f is differentiable
at a provided the limit in (4.3) exists. That’s the way differentiability is
1. DIFFERENTIABILITY 59
usually deﬁned in introductory calculus books. When the limit exists, it’s
indicated by f
(a) and called the value of the derivative of f at a.
While it’s natural to emphasize this limit when the goal is ﬁnding
derivatives, we’ve taken a somewhat different point of view to emphasize
a different aspect of the theory. Our goal is to develop properties of
functions known to be differentiable, so we’re adopting the deﬁnition in a
form that simpliﬁes their discovery and veriﬁcation. For example, when f
is differentiable at a, the way we’ve expressed f (x) in terms of continuous
functions makes it obvious that f is continuous at a.
Our deﬁnition of differentiability makes it easy to appreciate the reason
that derivatives are used to search for extreme values of functions. When
f is differentiable at a with f
(a) = 0, the continuity of g at a in the
representation
f (x) −f (a) = (x −a) g (x)
requires g (x) to have the same sign as f
(a) at all points in some open
interval (a −δ, a +δ), so in that interval the values of f (x) − f (a) are
positive on one side of a and negative on the other. So when f
(a) exists
and is nonzero, we see that f (a) is neither the largest nor the smallest value
of f (x). Points in the domain of f at which f
is either undeﬁned or zero
are called critical points; when we’re searching for extreme values in an
open interval we can conﬁne our search to critical points.
While our concept of differentiability really is the same as that given in
most introductory calculus books, a meaningful extension of the concept
is possible and is sometimes used. It isn’t entirely necessary to insist that
the domain of f include an open interval I about a; when the domain of f
is a closed interval some authors allow a to be an endpoint. We’ve chosen
not to allow that. Our point of view is that a function differentiable at a
should closely resemble a linear function near x = a, and linear functions
are deﬁned at all the points near a, not just some of them.
As we mentioned above, we generally indicate
lim
x→a
f (x) −f (a)
x −a
= f
(a).
We think of f
(a) as the value of a new function f
derived from f, called
the derivative of f. The function f
is not the same as the function g in our
deﬁnition of differentiability at the point a; note f
and g have the same
value at a but not necessarily at any other points. If f : E → R, then the
domain of f
is the set of points in E at which f is differentiable. We call
f a differentiable function if it is differentiable at all points in its domain,
60 CHAPTER IV THE DERIVATIVE
and in that case f and f
have the same domain. The operation we perform
on f to create the function f
is called differentiation; it is an operation that
we perform on functions, not on numbers. Consequently, two functions
that have the same value at x = a won’t necessarily have equal derivatives
at x = a. It is true that equal functions have equal derivatives, but equal
functions must have equal values at all points, not just at a.
When f is differentiable at all points of an interval, it’s natural to
expect f
to be continuous there since its values are found by removing
discontinuities. That happens to be true for many of the functions we work
with, but it’s false in general. The discontinuity we remove to ﬁnd f
(a)
isn’t a discontinuity of f
(x), it’s a discontinuity of another function, and
f
may well have a nonremovable discontinuity at points in its domain.
Nonetheless, when f
is deﬁned on an interval it does indeed have some
of the properties we associate with continuous functions. We’ll investigate
this point in the third section of this chapter.
It’s common to interpret differentiability geometrically in terms of the
existence of a line tangent to a curve. Without worrying too much about
when a set of points forms a curve, it is now agreed that a line L through
a point P on a curve C is a tangent line provided that whenever a moving
point Q on C approaches P, the angle between L and the line through P
and Qapproaches zero. The directions of nonvertical lines are determined
by their slopes, so we say that the nonvertical line L through P is the
tangent line at P if the slope of the line through P and Q approaches the
slope of L as the moving point Q approaches P along C. If our curve is
the graph of f and P is the point (a, f (a)), we may take Qto be (x, f (x))
with x = a. Then the slope of the line through P and Q is exactly
f (x) −f (a)
x −a
,
and when it has a limit f is differentiable at a.
There are, in fact, other ways that we can specify a curve without
giving y explicitly as a function of x, and it may still be possible to ﬁnd the
slope of a tangent line by taking a limit of slopes. If we think of (x, y) as
the ﬁxed point P on C and (x + ∆x, y + ∆y) as the moving point Q, then
the slope of the line through P and Q is ∆y/∆x, with ∆x and ∆y both
approaching zero as Q approaches P. The slope of the tangent line at P is
then
dy
dx
= lim
∆x→0
∆y
∆x
,
provided this limit exists.
EXERCISES 61
The symbol
dy
dx
for the slope of a tangent line was introduced by Got
tfried Leibniz, who discovered many of the basic principles of calculus at
about the same time as Newton. When y = f (x) with f a differentiable
function,
dy
dx
and f
(x) have exactly the same meaning and are used almost
interchangeably. However, there’s no convenient way to indicate the coor
dinates of the point of tangency in Leibniz’s notation. In some situations,
that’s a deﬁnite disadvantage. But the f
(x) notation can be used only
when some function named f has been deﬁned, and sometimes we ﬁnd the
derivative before we get around to introducing f. In such cases, Leibniz’s
notation is more convenient.
Our notation f
(x) for the derivative wasn’t Newton’s choice either;
in fact, he didn’t usually think of y as a function of x. Instead, he preferred
to think of (x, y) as the coordinates of a moving point, with both x and y
functions of time. Assuming that y = f (t) and x = g (t), we have
∆y = f (t + ∆t) −f (t) and ∆x = g (t + ∆t) −g (t).
Newton assumed that x and y were differentiable functions of t, with
derivatives
˙ x = f
(t) and ˙ y = g
(t).
As long as ˙ x = 0, the slope of the tangent line is
lim
∆t→0
∆y
∆x
= lim
∆t→0
∆y/∆t
∆x/∆t
=
˙ y
˙ x
,
which was Newton’s notation for the slope of the tangent line.
EXERCISES
1. Use the deﬁnition of differentiability to showthat if f is differentiable
at the point a, then so is f
2
.
2. By considering the cases a > 0, a = 0, and a < 0 separately, show
that f (x) = xx deﬁnes a function that is differentiable at each
a ∈ R.
3. Use the deﬁnition of differentiability to show that
f (x) =
_
x +x
2
sin(1/x), x = 0
0, x = 0
deﬁnes a function that is differentiable at 0, with f
(0) = 1.
Note: Everywhere except at the origin, the derivative of f is given by
f
(x) = 1 + 2xsin(1/x) −cos (1/x).
62 CHAPTER IV THE DERIVATIVE
The discontinuity of f
at 0 is not removable, even though f
(0)
exists.
4. Show that if (x, y) and (x + ∆x, y + ∆y) are points on the circle
x
2
+y
2
= r
2
with y, ∆x, and y + ∆y nonzero, then
∆y
∆x
= −
2x + ∆x
2y + ∆y
and
dy
dx
= −
x
y
.
2 COMBINING DIFFERENTIABLE FUNCTIONS
We’ve seen that when we combine continuous functions in a number of
ways, the new functions we produce are also continuous. We’d like to
be able to recognize that the same sorts of combinations of differentiable
functions are differentiable. We’ll see how that works by ﬁrst deriving
rules for the derivative of basic arithmetic combinations of differentiable
functions. Then we’ll give a formal statement and rigorous proof of the
chain rule. The formulas we derive in this section should be familiar from
basic calculus; what may be unfamiliar is the way the formulas follow
naturally from the deﬁnition of differentiable functions and derivatives.
Suppose we have two functions f
1
and f
2
that are known to be differ
entiable at the same point a, and that f is a newfunction deﬁned in terms of
f
1
and f
2
. First we note that there must be an open interval I
1
containing
a and a function g
1
that is continuous at a and satisﬁes
f
1
(x) = f
1
(a) + (x −a) g
1
(x) for all x ∈ I
1
. (4.4)
Also, there is an open interval I
2
containing a and a function g
2
that is
continuous at a and satisﬁes
f
2
(x) = f
2
(a) + (x −a) g
2
(x) for all x ∈ I
2
. (4.5)
Then I = I
1
∩ I
2
deﬁnes an open interval containing a on which both
equations are true. Our strategy is to use equations (4.4) and (4.5) to
express
f (x) = f (a) + (x −a) g (x) for all x ∈ I
with g (x) deﬁned in terms of f
1
(a), f
2
(a), g
1
(x), and g
2
(x). Whenever
we can recognize g as continuous at a, we’ll know that f is differentiable
at a with f
(a) = g (a).
First we consider linear combinations of f
1
and f
2
. For this case we
deﬁne f : I →Rby
f (x) = c
1
f
1
(x) +c
2
f
2
(x)
2. COMBINING DIFFERENTIABLE FUNCTIONS 63
with c
1
and c
2
unspeciﬁed constants. For all x ∈ I, we may write
f (x) = c
1
[f
1
(a) + (x −a) g
1
(x)] +c
2
[f
2
(a) + (x −a) g
2
(x)]
= c
1
f
1
(a) +c
2
f
2
(a) + (x −a) [c
1
g
1
(x) +c
2
g
2
(x)]
= f (a) + (x −a) g (x),
with g deﬁned by
g (x) = c
1
g
1
(x) +c
2
g
2
(x).
Since g is continuous at a whenever g
1
and g
2
are, we recognize that f is
differentiable at a. Moreover, since
g (a) = c
1
g
1
(a) +c
2
g
2
(a),
we’ve proved that
(c
1
f
1
+c
2
f
2
)
(a) = c
1
f
1
(a) +c
2
f
2
(a).
Next we consider the product of f
1
and f
2
. This time we deﬁne the
numerical function f on I by
f (x) = f
1
(x) f
2
(x).
Then for all x ∈ I we may write
f (x) =[f
1
(a) + (x −a) g
1
(x)] [f
2
(a) + (x −a) g
2
(x)]
=f
1
(a) f
2
(a) + (x −a) f
1
(a) g
2
(x)
+ (x −a) g
1
(x) f
2
(a) + (x −a)
2
g
1
(x) g
2
(x)
=f (a) + (x −a) g (x),
with g the numerical function on I deﬁned by
g (x) = f
1
(a) g
2
(x) +g
1
(x) f
2
(a) + (x −a) g
1
(x) g
2
(x).
Again we recognize g as continuous at a, proving that f is differentiable at
a. Since
g (a) = f
1
(a) g
2
(a) +g
1
(a) f
2
(a),
we’ve established the product rule
(f
1
f
2
)
(a) = f
1
(a) f
2
(a) +f
1
(a) f
2
(a).
64 CHAPTER IV THE DERIVATIVE
Now we consider dividing f
1
by f
2
. Since we can’t divide by zero, we
should assume that f
2
(a) = 0. Since f
2
is continuous at a, we can assume
that our open interval I
2
was chosen in such a way that f
2
(x) is never 0 on
I
2
, and then we deﬁne a numerical function f on I by
f (x) =
f
1
(x)
f
2
(x)
=
f
1
(a) + (x −a) g
1
(x)
f
2
(a) + (x −a) g
2
(x)
.
Regrouping this last expression in the form f (a) +(x −a) g (x) is a little
trickier here. We get
g (x) =
f (x) −f (a)
x −a
=
1
(x −a)
_
f
1
(a) + (x −a) g
1
(x)
f
2
(a) + (x −a) g
2
(x)
−
f
1
(a)
f
2
(a)
_
=
[f
2
(a) g
1
(x) −f
1
(a) g
2
(x)]
[f
2
(a) + (x −a) g
2
(x)] f
2
(a)
because the terms involving the product of f
1
(a) and f
2
(a) cancel out
when we subtract. Since f
2
(a) = 0, we again see that g is continuous at
a, proving that f is differentiable at a. Since
g (a) =
f
2
(a) g
1
(a) −f
1
(a) g
2
(a)
[f
2
(a)]
2
,
we’ve established the quotient rule
_
f
1
f
2
_
(a) =
f
2
(a) f
1
(a) −f
1
(a) f
2
(a)
[f
2
(a)]
2
.
That takes care of the rules for differentiating linear combinations,
products, or quotients of differentiable functions. We still need to treat the
composition of differentiable functions; that’s the subject of the chain rule.
We’ll need a new set of assumptions here, so we’ll state them formally.
Theorem 2.1: Suppose that f is a function that is differentiable
at the point a, and that F is a function that is differentiable at the point
f (a). Then there is an open interval I containing the point a on which the
composition
H (x) = F (f (x))
deﬁnes a function H that is differentiable at the point a. The value of its
derivative there is
H
(a) = F
(f (a)) f
(a).
2. COMBINING DIFFERENTIABLE FUNCTIONS 65
Proof : Since f is differentiable at a, we know there is an open
interval I
0
about a and a function g : I
0
→ R that is continuous at a, and
that satisﬁes
f (x) = f (a) + (x −a) g (x) for all x ∈ I
0
. (4.6)
Also, since F is differentiable at f (a), we know there is an open interval
I
1
about f (a) and a function G : I
1
→ R that is continuous at f (a), and
that satisﬁes
F (y) = F (f (a)) + [y −f (a)] G(y) for all y ∈ I
1
. (4.7)
Our plan is to substitute y = f (x) in (4.7) with f (x) given by (4.6), but
ﬁrst we need to choose the interval I in a way that makes this substitution
valid for all x ∈ I.
Since I
1
is an open interval about f (a), there must be an ε > 0 such
that (f (a) −ε, f (a) +ε) ⊂ I
1
, and since f is continuous at a, there must
be a δ > 0 such that
f (x) −f (a) < ε for all x ∈ I
0
∩ (a −δ, a +δ).
Then for each x ∈ I = I
0
∩(a −δ, a +δ), we get f (x) ∈ I
1
and therefore
H (x) = F (f (x)) = F (f (a)) + [f (x) −f (a)] G(f (x))
by (4.7). Since F (f (a)) = H (a) and f (x) satisﬁes (4.6), we see that
H (x) = H (a) + (x −a) g (x) G(f (x)) for all x ∈ I.
Since g (x) G(f (x)) is continuous at a, we see that H is differentiable at
a with
H
(a) = g (a) G(f (a)) = f
(a) F
(f (a)).
That completes our proof of the chain rule.
We should note that the chain rule gives sufﬁcient conditions for a
function of a function to be differentiable; the conditions are by no means
necessary. For example, f (x) =
3
√
x is deﬁned and continuous for all x,
but it is not differentiable at x = 0. Clearly,
f (x) −f (0)
x −0
=
x
1/3
x
= x
−2/3
66 CHAPTER IV THE DERIVATIVE
and x
−2/3
has an inﬁnite discontinuity at 0, not a removable one. Yet for
F (x) = x
3
,
H (x) = F (f (x)) =
_
3
√
x
_
3
= x for all x,
and so H is clearly differentiable at 0 with H
(0) = 1.
Sometimes it is easier to work with an identity involving the values
of a function than with an explicit formula for the function itself. In such
cases, we may be able to express f (x) in the form f (a) + (x −a) g (x)
by deﬁning g in terms of f rather than giving a complete formula for it.
In these cases, the continuity of f becomes an important part of showing
that g is continuous. We will see systematic ways of exploiting this idea in
Sections 4.6 and 4.7, and it is the key to Exercise 8 below.
EXERCISES
5. According to the chain rule, if F (x) = f
_
x
2
_
and f is differentiable
at a
2
, then F is differentiable at a with F
(a) = 2af
_
a
2
_
. Use the
deﬁnition to prove this directly.
6. If the numerical function f is differentiable at a and F (x) = 1/f (x),
then F
(a) = −f
(a) /f (a)
2
provided that f (a) = 0. Show that
this follows from the quotient rule and also from the chain rule.
7. Suppose that f and F are functions such that F (f (x)) = x for all
x in an open interval I. Show that if a is any point in I such that
f is differentiable at a and F is differentiable at f (a), then neither
F
(f (a)) nor f
(a) is 0.
8. Suppose that F (f (x)) deﬁnes a function that is differentiable at a.
Prove that if f is continuous at a and f (a) is not a critical point for
F, then f is differentiable at a.
3 MEAN VALUES
In formal mathematics, meanis often used instead of the more familiar term
average, probably to alert the reader that a precisely deﬁned mathematical
concept is intended rather than the loose idea conveyed by phrases such as
“average person” or “average day.” When we talk about the mean value
of a function, it’s important to realize that we’re referring to the mean of a
set of values of f, not necessarily to a value of the function. We’ve seen
a similar distinction in our study of the intermediate value theorem, where
the goal was to prove that each number intermediate between two values
of a continuous function is also a value of the function.
3. MEAN VALUES 67
When f is deﬁned on [a, b], the quantity
m =
f (b) −f (a)
b −a
is sometimes thought of as representing the mean value of the derivative
of f over [a, b]. That’s because the net change of f over [a, b] is the same
as that for mx, and the derivative of mx is always equal to m. This is
analogous to the case when the interval [a, b] represents a time interval and
the values of f represent position. In that case, values of f
(t) are called
instantaneous velocities and m is called the average velocity over the time
interval.
Clearly some hypotheses must be satisﬁed before we can assert that
the mean value of a function is actually one of its values. It’s reasonable
to expect that continuity is involved somehow since we needed that as
sumption to prove the intermediate value theorem. The theorem below,
universally known as the mean value theorem, gives conditions under
which the mean value of a derivative is actually a value of the derivative.
Note that the hypotheses are surprisingly weak; continuity of the derivative
need not be assumed. However, the function itself must be continuous.
Theorem 3.1: Let f be a numerical function that is continuous
on [a, b] and differentiable at every point of (a, b). Then there is a point
ξ ∈ (a, b) such that
f
(ξ) =
f (b) −f (a)
b −a
.
The mean value theorem is both deep and powerful. Many newer
calculus books are attempting to leave it out. There are indeed more
intuitive explanations available for many of its traditional applications;
none, however, can rival the mean value theorem for simplicity and power.
We’ll use it in just about every remaining section of this book.
Proof : We’ll follow the route taken by most calculus texts that
include the mean value theorem, and justify it by appealing to an interme
diate result, generally called Rolle’s theorem. Rolle’s theorem states that
if a < b and f is continuous on [a, b] with f (a) = f (b), then f has a
critical point in (a, b). To prove it, we observe that for f continuous on
[a, b], the extreme value theorem tell us that f has both a least and greatest
value over that interval. We’ve assumed that f (a) = f (b), so f must have
at least one of its extreme values at an interior point of [a, b], and such an
interior point must be a critical point.
68 CHAPTER IV THE DERIVATIVE
It’s a simple step from Rolle’s theorem to the mean value theorem. If
f is continuous on [a, b] and we deﬁne F : [a, b] →Rby
F (x) = (b −a) f (x) −(x −a) [f (b) −f (a)],
then clearly F is continuous and
F (b) = (b −a) f (b) −(b −a) [f (b) −f (a)] = (b −a) f (a) = F (a).
So Rolle’s theoremguarantees that F has a critical point ξ ∈ (a, b). Clearly,
F is differentiable everywhere that f is, with
F
(x) = (b −a) f
(x) −[f (b) −f (a)].
Since f has been assumed to be differentiable on (a, b), F is differentiable
at its critical point ξ, and so
0 = F
(ξ) = (b −a) f
(ξ) −[f (b) −f (a)].
Solving for f
(ξ) completes the proof.
The simplicity of these arguments raises an obvious question: what’s
so deep about the mean value theorem? Certainly there’s nothing at all
remarkable about the step fromRolle’s theoremto the mean value theorem.
The difﬁculty is in proving Rolle’s theorem, and the only place where that
argument is not routine is in establishing the existence of a point ξ in (a, b)
where f has an extreme value. That’s what makes the mean value theorem
a deep result. When we proved that every function continuous on a closed
interval had both a greatest and a least value there, we were unable to
give a general method for locating the points where they occurred. The
statement of the mean value theorem gives no hint as to how the point ξ
can be located, nor does the proof we gave provide any clues to ﬁnding it.
There are proofs of the mean value theoremthat make no use of extreme
values, and the problem of locating ξ is simpler than locating an extreme
value. But we won’t pursue the search for ξ, since our ability to locate it
doesn’t add to the usefulness of the mean value theorem. In fact, when
we do ﬁnd ξ we don’t need the mean value theorem; a theorem is never
needed to demonstrate the existence of something already at hand. The
point of the theorem is that we can use the range of values of f
over (a, b)
to estimate f (b) −f (a) as a multiple of b −a. It may not be necessary to
knowf (a) and f (b) or even a and b to do so. The simplest case is when f
is 0 everywhere. In that case, we conclude that f must be constant. When
f
keeps the same sign throughout (a, b), the mean value theorem tells us
that f (b) −f (a) shares that sign.
3. MEAN VALUES 69
One remarkable consequence of the mean value theorem is that when
f is differentiable on an interval I, the set of values {f
(x) : x ∈ I} is also
an interval, even when f
has discontinuities in I. To prove this, suppose
that [a, b] ⊂ I, and that r is any number strictly between f
(a) and f
(b).
Since f must be continuous on I, for 0 < h < b −a the formula
s
h
(x) =
f (x +h) −f (x)
h
deﬁnes a continuous function s
h
on [a, b −h]. Since
lim
h→0
s
h
(a) = f
(a) and lim
h→0
s
h
(b −h) = f
(b),
if we take h small enough then r will be between s
h
(a) and s
h
(b −h) as
well as between f
(a) and f
(b). For such an h, the intermediate value
theorem guarantees there is a c ∈ (a, b −h) with
r = s
h
(c) =
f (c +h) −f (c)
h
.
Then the mean value theorem guarantees there is a ξ ∈ (c, c +h) ⊂ (a, b)
with
f
(ξ) =
f (c +h) −f (c)
h
= r.
The next theorem illustrates a more typical use of the mean value
theorem. It shows that derivatives must have another property that we
generally associate with continuous functions, even though derivatives are
allowed to have discontinuities.
Theorem 3.2: Let a be a point in a given open interval I, and let
f be a continuous function on I. If f is known to be differentiable at all
points of I except possibly at a and lim
x→a
f
(x) exists, then f must also
be differentiable at a with
f
(a) = lim
x→a
f
(x).
Proof : We deﬁne g (x) = [f (x) −f (a)] / (x −a) for x ∈ I but
x = a, and then note that the conclusion of the theorem is simply that
lim
x→a
g (x) = lim
x→a
f
(x).
Calling L the limit of f
, we need to show that when ε > 0 is given, there
must be a δ > 0 such that every x ∈ (a −δ, a +δ) except possibly a itself
70 CHAPTER IV THE DERIVATIVE
satisﬁes g (x) −L < ε. We know there is a δ such that f
(x) −L <
ε for all x = a in (a −δ, a +δ), but f
and g are different functions.
Fortunately, their values are strongly related. For x ∈ I but x = a, we
can apply the mean value theorem to f on [x, a] when x < a and on [a, x]
when a < x; in either case, we see there must be a point ξ between x and
a with
f
(ξ) =
f (a) −f (x)
a −x
=
f (x) −f (a)
x −a
= g (x).
So if all the values of f
on (a −δ, a +δ) are between L − ε and L + ε,
the same must be true of the values of g.
We should note that the mean value theorem only works one direction.
For a given ξ, the value f
(ξ) may not be the slope of any line segment
joining two points on the graph. The simplest example involves f (x) = x
3
;
obviously f
(0) = 0. However, for a = b we see
f (b) −f (a)
b −a
= b
2
+ab +a
2
=
1
2
_
a
2
+b
2
+ (a +b)
2
_
,
and that must be positive. Theorem 3.2 only works one way for the same
reason, so the formula for f
(x) when x = a isn’t always useful for ﬁnding
f
(a).
Here’s another theorem that can be very useful. It’s sometimes called
Cauchy’s mean value theorem, in honor of the French mathematician Au
gustin Cauchy (1789–1857). Cauchy’s name is associated with many
fundamental concepts in the mathematics that developed from calculus,
including the notion of limits. One of the uses of this theorem is to prove
L’Hˆ opital’s rule, a method for using derivatives to ﬁnd limits instead of
using limits to ﬁnd derivatives. We’ll develop L’Hˆ opital’s rule in Chapter
9.
Theorem 3.3: Suppose that f and g are continuous on [a, b] and
differentiable on (a, b), with g
never vanishing there. Then there is a point
ξ ∈ (a, b) such that
f (b) −f (a)
g (b) −g (a)
=
f
(ξ)
g
(ξ)
.
Proof : Before we begin, we note that the mean value theorem
implies that g (b) − g (a) = 0 since g
(x) = 0 for a < x < b. Thus the
conclusion of the theorem does not involve division by zero.
EXERCISES 71
Cauchy’s mean value theorem is also a simple consequence of Rolle’s
theorem. This time we deﬁne the auxiliary function F : [a, b] →Rby the
formula
F (x) = [f (b) −f (a)] g (x) −[g (b) −g (a)] f (x).
It’s obviously continuous on [a, b] and differentiable on (a, b), so any
critical point for F in (a, b) must be a point where F
vanishes. When we
evaluate either F (a) or F (b), some terms cancel and we ﬁnd
F (a) = f (b) g (a) −f (a) g (b) = F (b).
Hence Rolle’s theorem guarantees that there is a point ξ ∈ (a, b) with
0 = F
(ξ) = [f (b) −f (a)] g
(ξ) −[g (b) −g (a)] f
(ξ).
Since neither g (b) −g (a) nor g
(ξ) is zero, this equation is equivalent to
the statement of the theorem.
EXERCISES
9. Use the mean value theorem to prove that for a < b,
¸
¸
¸
_
1 +b
2
−
_
1 +a
2
¸
¸
¸ < b −a.
10. Use the mean value theorem to prove that if 0 ≤ a < b then
3a
2
(b −a) < b
3
−a
3
< 3b
2
(b −a).
Why is the restriction 0 ≤ a necessary?
11. Suppose that f and g are positive, differentiable functions on an
interval I, with
f
(x)
f (x)
<
g
(x)
g (x)
for all x ∈ I.
Prove that F (x) = f (x) /g (x) deﬁnes a decreasing function on I.
12. Under the hypotheses of Theorem 3.3, show that
lim
x→a+
f (x) −f (a)
g (x) −g (a)
= lim
x→a+
f
(x)
g
(x)
when the second limit exists. This is one of the arguments used to
establish L’Hˆ opital’s rule.
72 CHAPTER IV THE DERIVATIVE
13. By considering the intervals between successive roots, prove that if
p (x) is a nonzero polynomial that vanishes at n different points, then
there must be at least n − 1 points where p
(x) vanishes. Then use
induction to prove that no polynomial of degree n can have more than
n roots.
4 SECOND DERIVATIVES AND APPROXIMATIONS
When the numerical function f is differentiable at the point a, we know
that its domain includes an open interval I about a, and that the function
g : I →Rdeﬁned by the formula
g (x) =
_
_
_
f (x) −f (a)
x −a
, x = a
f
(a), x = a
is continuous at a. If g is itself differentiable at a, then there must be
another function h : I →Rthat is continuous at a and satisﬁes
g (x) = f
(a) + (x −a) h(x) for all x ∈ I.
For x ∈ I, we can combine the formulas for f (x) and g (x):
f (x) = f (a) + (x −a) g (x)
= f (a) + (x −a) f
(a) + (x −a)
2
h(x).
Rewriting
h(x) = h(a) + [h(x) −h(a)],
we obtain
f (x) =f (a) + (x −a) f
(a) + (x −a)
2
h(a)
+ (x −a)
2
[h(x) −h(a)].
The ﬁrst three terms in this last expression deﬁne a simple quadratic
polynomial, and the last term represents the error when we use this polyno
mial to approximate f (x). Since his continuous at a, the last termbecomes
smaller than any nonzero constant multiple of (x −a)
2
as x approaches
a. So we generally expect this quadratic approximation to be much more
accurate than the linearization of f (x) near x = a. Quadratic functions
are fairly easy to work with, so such an approximation is potentially quite
useful. To make quadratic approximations easy to use, we need a simple
way to verify that g is differentiable at a and to calculate g
(a).
4. SECOND DERIVATIVES AND APPROXIMATIONS 73
Since g (a) = f
(a), for all x ∈ I except a we must have
h(x) =
f (x) −f (a) −(x −a) f
(a)
(x −a)
2
,
an expression that clearly has a discontinuity at x = a. We can remove the
discontinuity when f satisﬁes some very simple conditions. If we assume
that f is differentiable not just at a but everywhere between a and x, and
that it is continuous at x as well as at a and at all the points between a
and x, then we can simplify the formula for h(x) with the aid of Cauchy’s
mean value theorem. We think of h as given by
h(x) =
F (x) −F (a)
G(x) −G(a)
with
F (x) = f (x) −f (a) −(x −a) f
(a) and G(x) = (x −a)
2
.
Note that F (a) = G(a) = 0. Since G
(x) = 2 (x −a) = 0 for x = a,
there must be a point ξ between a and x such that
h(x) =
F
(ξ)
G
(ξ)
=
f
(ξ) −f
(a)
2 (ξ −a)
. (4.8)
It doesn’t really matter whether x > a or x < a. In the ﬁrst case we use the
interval [a, x] and in the second we use [x, a]. We cover both possibilities
by simply saying that ξ is between a and x.
When f
is differentiable at the point a, we can remove the discontinuity
in h at a by deﬁning
h(a) = lim
x→a
h(x) = lim
ξ→a
f
(ξ) −f
(a)
2 (ξ −a)
=
1
2
f
(a).
The quantity f
(a) is called the second derivative of f at a, and it
represents the value at a of the derivative of f
. Returning to (4.8), when
f
is differentiable everywhere between a and x, we can apply the mean
value theorem to obtain
h(x) =
1
2
f
(ξ
∗
)
for some point ξ
∗
between a and ξ and therefore between a and x.
Now let’s assemble all these ideas in a formal theorem, just to make
sure we include all the needed hypotheses. We’ll use it in Chapter 6 to
74 CHAPTER IV THE DERIVATIVE
analyze a limit that is sometimes used to deﬁne the natural exponential
function, and we’ll use it again in Chapter 9 to analyze Newton’s method
for solving nonlinear equations.
Theorem 4.1: Suppose that the numerical function f is continuous
on [α, β] and differentiable in (α, β) and that f
is differentiable at the
point a ∈ (α, β). Then there is a continuous function h : [α, β] → Rwith
h(a) =
1
2
f
(a) such that
f (x) = f (a) + (x −a) f
(a) + (x −a)
2
h(x) for all x ∈ [α, β].
With the additional assumption that f
is differentiable everywhere in (α, β),
for each x = a in [α, β] there is a point ξ between a and x satisfying
f (x) = f (a) + (x −a) f
(a) +
1
2
(x −a)
2
f
(ξ).
Assuming only the existence of f
(a) gives us a qualitative appraisal
of the quadratic approximation
f (a) + (x −a) f
(a) +
1
2
(x −a)
2
f
(a)
to f (x) near x = a. This approximation differs from f (x) by an amount
that becomes arbitrarily small in comparison to (x −a)
2
as x approaches
a. When f
(x) exists throughout (α, β), we have a convenient quantitative
expression for this difference in the form
1
2
(x −a)
2
_
f
(a) −f
(ξ)
¸
for some ξ between a and x.
While quadratic approximations are often used in quantitative work,
the theorem also has important qualitative implications. The continuity of
h at a shows that h(x) and f
(a) must have the same sign at points near
a, so the graph of f is above the tangent line where f
> 0 and below
the tangent line where f
< 0. In particular, where the tangent line is
horizontal the sign of the second derivative determines whether the point
of tangency represents a relative maximum or relative minimum; this is the
familiar second derivative test for relative extrema.
There’s a more general version of Theorem 4.1 that deals with approx
imations using polynomials of higher degree with hypotheses involving
derivatives of higher order. It’s known as Taylor’s theorem, and we dis
cuss it in Chapter 8. Such higherorder approximations are potentially
more accurate but somewhat harder to work with.
EXERCISES 75
Figure 4.1 is a graph of f (x) = 1/
_
x
2
+ 1
_
along with its linearization
and quadratic approximation around x = 1, with the graph of f shown as
a heavy curve. Near x = 1 the quadratic approximation produces a curve
that almost exactly matches the original graph, making it a noticeably better
approximation than the linearization; however, at more distant points the
quadratic approximation may be less accurate.
0
0.2
0.4
0.6
0.8
1
1.2
0.5 1 1.5 2
Figure 4.1 Linearization and quadratic approximation for 1/
_
x
2
+ 1
_
.
EXERCISES
14. Prove that if f
(a) exists and is not zero, then there is an open interval
I about a where the graph of f falls inside the graph of the parabola
f (a) + (x −a) f
(a) +
1
4
(x −a)
2
f
(a).
15. The quadratic approximation to
√
x near x = 4 is
Q(x) = 2 +
1
4
(x −4) −
1
64
(x −4)
2
.
Find a representation for
√
x−Q(x), and use it to bound the error in
the quadratic approximation by a small constant multiple of (x −4)
2
for x ∈ [3, 5].
5 HIGHER DERIVATIVES
Since the derivative of a numerical function is another numerical function,
as long as we continue to produce differentiable functions we can repeat the
76 CHAPTER IV THE DERIVATIVE
process of differentiation indeﬁnitely. Instead of using additional primes
to indicate higher derivatives, we write f
(n)
for the function obtained
from f by differentiating it n times. To unify our notation, we agree
that f
(0)
is simply f. Leibniz’s notation for the nth derivative is
d
n
y
dx
n
;
we’ll occasionally ﬁnd this notation more convenient. Mathematicians call
C
n
(I) the class of all numerical functions such that f
(n)
is continuous
on I. Since every point where a function is differentiable is a point of
continuity for the function, we see that the classes C
n
(I) decrease with n;
that is, C
n+1
(I) ⊂ C
n
(I) for all natural numbers n. Functions that are in
C
n
(I) for every natural number n are said to be inﬁnitely differentiable
on I; C
∞
(I) represents this class of functions. Every polynomial function
is inﬁnitely differentiable at all points, but the class C
∞
(I) includes many
other functions even when I is all of R.
Higher derivatives are useful in trying to ﬁnd the roots of polynomials,
a problem that comes up in many contexts. We often try to ﬁnd roots of
a polynomial p by simply seeing whether or not p (x) = 0 for various
values of x. Repeated roots, however, present special difﬁculties. Clearly,
something else besides evaluating p (x) must be done to identify a repeated
root. Derivatives of higher order can come to the rescue, and it’s easy to
understand why.
When x = a is a root of the polynomial p (x), we call it a kfold root
if k > 0 and there is a polynomial q (x) such that
p (x) = (x −a)
k
q (x) with q (a) = 0.
In this case, the product rule tells us that
p
(x) = k (x −a)
k−1
q (x) + (x −a)
k
q
(x)
= (x −a)
k−1
_
kq (x) + (x −a) q
(x)
¸
= (x −a)
k−1
q
1
(x),
where
q
1
(x) = kq (x) + (x −a) q
(x).
Note that q
1
is also a polynomial, and if k > 1 then p
(x) has a (k −1)fold
root since
q
1
(a) = kq (a) = 0.
Taking additional derivatives shows us that for n ≤ k,
p
(n)
(x) = (x −a)
k−n
q
n
(x),
5. HIGHER DERIVATIVES 77
where q
n
(x) is a polynomial with q
n
(a) = 0.
So when p has a kfold root at a, then p
(n)
(a) = 0 for n < k but not
for n = k. That’s a theorem and we’ve just proved it; we’ll state it next.
Theorem 5.1: Let p (x) be a polynomial and let k be a positive
integer. Then p (x) has a kfold root at x = a if and only if
p (a) = p
(a) = p
(a) = · · · = p
(k−1)
(a) = 0 with p
(k)
(a) = 0.
As we use more sophisticated applications for derivatives, we encounter
situations where we need to take an nth derivative of the product of two
functions. Fortunately, there’s a fairly simple formula that covers such
cases, known as Leibniz’s rule. Here’s a statement of it.
Theorem 5.2: Suppose that x is a point at which both f and g can
be differentiated n times. Then their product h can also be differentiated n
times at x, and
h
(n)
(x) =
n
k=0
n!
k! (n −k)!
f
(n−k)
(x) g
(k)
(x).
Proof : We prove the theorem using mathematical induction. The
case n = 1 is simple; it says that when f and g are differentiable at x their
product h is also differentiable, and
h
(1)
(x) =
1!
0!1!
f
(1)
(x) g
(0)
(x) +
1!
1!0!
f
(0)
(x) g
(1)
(x).
That’s just a messy statement of the familiar product rule, so we know that
Leibniz’s rule is true when n = 1.
Now we assume that N is a positive integer such that Leibniz’s rule is
true for n = N, and prove that it must also be true for n = N + 1. The
hypotheses for this case are that f and g may each be differentiated N +1
times at x. Then they’re also differentiable N times at x, and since we’ve
assumed that Leibniz’s rule is true for n = N,
h
(N)
(x) =
N
k=0
N!
k! (N −k)!
f
(N−k)
(x) g
(k)
(x).
When we examine the orders of the derivatives of f and g appearing in
this sum, we note that both k and N − k are less than N + 1. So by
the product rule, each term f
(N−k)
(x) g
(k)
(x) is also differentiable, and
78 CHAPTER IV THE DERIVATIVE
hence h
(N+1)
(x) is given by
N
k=0
N!
k! (N −k)!
_
f
(N+1−k)
(x) g
(k)
(x) +f
(N−k)
(x) g
(k+1)
(x)
_
.
To complete the proof, we need to show that this sum can be regrouped
as
N+1
k=0
(N + 1)!
k! (N + 1 −k)!
f
(N+1−k)
(x) g
(k)
(x).
That’s possible with a diligent application of algebra, but there’s an easier
way that’s also more informative. According to the binomial theorem,
(A+B)
N
=
N
k=0
N!
k! (N −k)!
A
N−k
B
k
,
so
(A+B)
N+1
= (A+B)
N
k=0
N!
k! (N −k)!
A
N−k
B
k
=
N
k=0
N!
k! (N −k)!
_
A
N+1−k
B
k
+A
N−k
B
k+1
_
.
However, we know that can be regrouped to give
(A+B)
N+1
=
N+1
k=0
(N + 1)!
k! (N + 1 −k)!
A
N+1−k
B
k
.
The algebra involved must be exactly the same for both cases, so no further
proof is necessary.
As we use Leibniz’s rule, it is often helpful to remember that the
numerical coefﬁcients are just the binomial coefﬁcients, and they can be
calculated easily fromPascal’s triangle. Indeed, Pascal’s triangle is nothing
more than an efﬁcient bookkeeping device for the cumulative effect of the
algebraic regrouping operation involved.
EXERCISES
16. Show that if g ∈ C
n
(I) and h(x) = xg (x), then h ∈ C
n
(I) and
h
(n)
(x) = xg
(n)
(x) +ng
(n−1)
(x).
6. INVERSE FUNCTIONS 79
(While this is a straightforward application of Leibniz’s rule, it is also
easy to establish by induction.)
17. Show that if g ∈ C
n
(I) and h(x) = x
2
g (x), then h ∈ C
n
(I) and
h
(n)
(x) = x
2
g
(n)
(x) + 2nxg
(n−1)
(x) +n(n −1) g
(n−2)
(x),
with the last term understood to be 0 when n = 1. Either Leibniz’s
rule or Exercise 16 can be used.
6 INVERSE FUNCTIONS
We often deal with pairs of functions having the property that one undoes
the other. For example, if f and φ are numerical functions deﬁned by the
formulas
f (x) = 2x + 1 and φ(x) =
1
2
(x −1),
then it’s easy to calculate that
f (φ(x)) = 2
_
1
2
(x −1)
_
+ 1 = x
as well as
φ(f (x)) =
1
2
([2x + 1] −1) = x.
We describe this situation by saying that f and φ are inverse functions;
we may either identify f as the inverse of φ or φ as the inverse of f.
In the example above, the formula for φwas calculated fromthe formula
for f. The procedure was to solve the equation
x = f (y)
for y in terms of x. The formula obtained for y was then used as the
formula for φ(x). This example was especially simple because we could
solve explicitly for y using basic algebra. For each x we found exactly one
y, and since the domain of f was all of Rwe knew that the y we found was
acceptable. In other cases we may not be so lucky; we may not know how
to perform the operations involved, and we may not be able to determine
the domains of the functions so easily. If we’re going to reason with inverse
functions correctly, we’ll need a more precise deﬁnition to deal with the
general case.
80 CHAPTER IV THE DERIVATIVE
Definition 6.1: Given two functions f : A → B and φ : B → A,
we say that f and φ are inverse functions if
φ(f (x)) = x for all x ∈ A and f (φ(y)) = y for all y ∈ B.
This is one context where we need to pay careful attention to all the
formalism in the language of functions. The question of whether f has an
inverse depends on both the sets A and B as well as the formula for f. For
example, if we deﬁne
f : [0, ∞) → [0, ∞) by f (x) = x
2
and
φ : [0, ∞) → [0, ∞) by φ(x) =
√
x,
then f and φ are inverse functions. But it we enlarge either the domain or
the range of f to include some negative numbers as well, then f no longer
has an inverse.
It’s important to note that the deﬁnition of inverse functions involves
two equations; it’s possible to satisfy one but not the other. For example,
suppose A = [0, 2], B = [0, 1], and we deﬁne f : A → B by f (x) =
x −1 and φ : B → A by φ(y) = 1 +y. While
f (φ(y)) = y = y for all y ∈ B
we note that
φ(f (x)) = 1 +x −1.
This only agrees with x when 1 ≤ x ≤ 2, not for all x ∈ A.
You may have noticed that something seems wrong with the example
we just considered. While φ : B → A we see that {φ(x) : x ∈ B},
the set of values of φ, is only [1, 2] instead of being all of A. That’s a
feature common to any example we might create. When f : A → B and
φ : B → A with f (φ(y)) = y for all y ∈ B, then φ(f (x)) = x for
each x ∈ A that is a value of φ. That’s easy to see; when x = φ(y) our
assumptions tell us that
φ(f (x)) = φ(f (φ(y))) = φ(y) = x.
Inverse functions are extremely convenient to work with when we have
them. For example, for each y ∈ B we can simply say that φ(y) is the
6. INVERSE FUNCTIONS 81
one value of x ∈ A such that f (x) = y. They’re often a convenient way
to identify a needed solution of an equation when we can’t or don’t want
to ﬁnd an explicit formula for it. The problem is that inverse functions are
elusive. Most functions don’t have inverses and algebra isn’t much help in
identifying many of those that do. Fortunately, the methods of calculus can
provide both simple ways for determining that an inverse does exist and
explicit methods for working with it, whether or not we ﬁnd its formula.
The theorem below shows how it’s done; we’ll naturally refer to it as the
inverse function theorem.
Theorem 6.1: Let I be an open interval, and let f be a numerical
function with domain I. If f has no critical points in I, then the image of
I under f is an open interval J, and f has an inverse function φ : J → I.
Moreover, φ is differentiable, and its derivative satisﬁes the equation
φ
(y) =
1
f
(x)
for all x ∈ I, y ∈ J with y = f (x).
The hypothesis that f has no critical points is a powerful one. Since
a critical point is any x in the domain of the function where either f
(x)
is undeﬁned or f
(x) = 0, we’ve assumed that f is differentiable at every
point in I and its derivative is never 0. While exploring the mean value
theorem, we saw that the set of values
_
f
(x) : x ∈ I
_
is itself an interval when f is differentiable at all points of the open interval
I, so to satisfy the hypothesis f
(x) must either be positive at all points in
I or be negative everywhere.
When we know that f and φ are inverse functions and that both are
differentiable, the relationship between their derivatives is an easy conse
quence of the chain rule. Since φ(f (x)) = x for all x ∈ I, we know
that
d
dx
(x) =
d
dx
{φ(f (x))} = φ
(f (x)) f
(x),
and we just substitute y for f (x) and recognize that the derivative of x is 1
to complete the derivation. But the theorem says more than that; we don’t
need to know in advance that both functions are differentiable to use it.
Proof : Now let’s prove the theorem. The ﬁrst assertion is that
the set of values
J = {f (x) : x ∈ I}
82 CHAPTER IV THE DERIVATIVE
deﬁnes an open interval. Since every differentiable function is continuous,
the intermediate value theoremguarantees that J must be an interval. Since
f has no critical points, none of its values can be extreme values. This
proves that J has neither a least or a greatest value, so J is an open interval.
The next thing to prove is that f has an inverse function. There’s only
one way possible to deﬁne it: for y ∈ J we must deﬁne φ(y) to be the one
x ∈ I such that f (x) = y. Since J was deﬁned to be the set of values of
f, we know there is at least one such x. If there were two, then Rolle’s
theorem would require that there be a critical point for f between them.
But we’ve assumed that there are no critical points in I, so for each y ∈ J
there is indeed exactly one x ∈ I with f (x) = y.
Instead of next proving that φ is differentiable, we’ll show that it’s
continuous on J; this turns out to be a necessary step in the argument.
Suppose that b ∈ J and we’re given an ε > 0. We need to ﬁnd a δ > 0
such that
φ(y) −φ(b) < ε for every y ∈ (b −δ, b +δ) ∩ J.
Call φ(b) = a ∈ I. Since I is an open interval, it doesn’t hurt to assume
that our given ε is small enough that both a −ε and a +ε are in I. Then
{f (x) : a −ε < x < a +ε}
is an open interval for exactly the same reasons that J is, and we know that
b is a point in it. We simply pick our δ small enough that
(b −δ, b +δ) ⊂ {f (x) : a −ε < x < a +ε} ,
and then every y ∈ (b −δ, b +δ) will satisfy
φ(y) ∈ (a −ε, a +ε).
Now we’re ready to show that φ is differentiable. In the process, we’ll
verify the formula for its derivative. We prove differentiability one point at
a time by analyzing φ(y) near each ﬁxed b ∈ J; the continuity of φ plays
an important role. Once again we’ll call
a = φ(b) and x = φ(y).
Since f is assumed to be differentiable at a, there is a numerical function
g deﬁned on I that is continuous at a and satisﬁes
f (x) = f (a) + (x −a) g (x) . (4.9)
EXERCISES 83
We can rewrite equation (4.9) as
y = b + [φ(y) −φ(b)] g (φ(y)),
and so as long as g (φ(y)) = 0 we have
φ(y) = φ(b) +
y −b
g (φ(y))
.
In fact, the function g deﬁned by equation (4.9) is never 0 on I, because
all its values are values of f
on I. That’s true by deﬁnition when x = a and
it’s a consequence of the mean value theorem for all other x ∈ I. Hence
the last equation is true for every y ∈ J, and the formula
G(y) =
1
g (φ(y))
deﬁnes a numerical function G on I. Since φ is continuous at b and g is
continuous at a, we see that Gis continuous at b. Hence φ is differentiable
at b, with
φ
(b) = G(b) =
1
g (φ(b))
=
1
g (a)
.
That completes our proof of the inverse function theorem.
EXERCISES
18. Prove that if f : A → B and φ : B → A with f (φ(y)) = y for all
y ∈ B, then f and φ are inverse functions unless there are x
1
, x
2
∈ A
with x
1
= x
2
but f (x
1
) = f (x
2
).
19. Show that f (x) = x
3
+ x deﬁnes a function on all of R that has a
differentiable inverse. Find the domain of the inverse function φ, and
calculate φ
(2).
20. Suppose f (x) =
1
2
_
−b +
√
b
2
−4c + 4x
_
, where b and c are ﬁxed
real numbers, and the domain of f is understood to be the set of all x
at which the given expression deﬁnes a differentiable function. Show
that f satisﬁes the hypotheses of Theorem 6.1, and ﬁnd an explicit
formula for its inverse φ. Calculate f
and φ
from their formulas,
and show they satisfy the conclusion of Theorem 6.1.
21. Suppose that f satisﬁes all the hypotheses of Theorem 6.1, and that
b = f (a) is a point in the domain of the inverse function φ. Prove
84 CHAPTER IV THE DERIVATIVE
that if f
is differentiable at a, then φ
is differentiable at b, and ﬁnd
a formula for φ
(b).
22. Even though f (x) = x
3
has a critical point at 0, f still has an inverse
with domain and range all of R; the inverse function is x
1/3
. How
can Theorem 6.1 be used to show that x
1/3
is differentiable at all
points except the origin?
7 IMPLICIT FUNCTIONS AND IMPLICIT
DIFFERENTIATION
Sometimes we need to work with the graph of an equation
F (x, y) = C
that we can’t solve either for y as a function of x or for x as a function
of y. Plotting such graphs is difﬁcult even for computers. Instead of
ﬁnding points by substituting values for one coordinate and calculating the
corresponding value of the other, it may be necessary to substitute possible
values for both coordinates and see whether the equation is satisﬁed. To
aid in the study of such graphs, most introductory calculus texts present
a technique known as implicit differentiation. To get started, we need
the coordinates of a point (a, b) on the graph. Brieﬂy, the method is to
assume that near the known point, the equation deﬁnes y as a differentiable
function φ(x) on an open interval I about a, in such a way that
F (x, φ(x)) = C for all x ∈ I.
Then φ
(x) can be computed by differentiating this equation with respect to
x and solving for φ
(x) in terms of x and φ(x). The equation to be solved
should be linear in the unknown φ
(x), and the procedure is supposed to
work as long as the coefﬁcient of φ
(x) is nonzero. Setting x = a and
φ(x) = b in the formula for φ
(x) then gives the slope of the tangent line
at (a, b).
Often the introduction of this method brings a great deal of confusion,
partly because the underlying concepts haven’t even been given names, let
alone clear descriptions. Lucid explanations of the method require some
understanding of functions of several variables and of the basic methods
for dealing with them, and that understanding is generally missing in basic
calculus courses. The study of functions of more than one variable gets
much more complicated than what we’ve been doing up to now. However,
the part of the theory needed to understand implicit differentiation is not
especially complicated; we’ll develop it below.
7. IMPLICIT FUNCTIONS AND IMPLICIT DIFFERENTIATION 85
Typically F (x, y) is a simple algebraic expression involving two un
knowns x and y. But in principle it can be anything that assigns a numerical
value to each point (x, y) in some region of the coordinate plane, as long as
the value assigned depends only on the location of the point. For (a, b) in
the region of the plane where F is deﬁned, we say that F is continuous at
(a, b) if for every given ε > 0, there is a number δ > 0 such that every point
(x, y) in that region with both x −a < δ and y −b < δ also satisﬁes
F (x, y) −F (a, b) < ε. This is exactly the same sort of condition we
developed for doing arithmetic with numbers known only approximately,
and we can use it in exactly the same way. It closely parallels the concept
of continuity we developed for functions deﬁned on subsets of the real line;
instead of requiring that a condition hold for all x in the interval of length
2δ centered at a, we require it to be true for all (x, y) inside the square of
side 2δ centered at (a, b).
Calculus teaches us to study functions by looking carefully at the
way their values change, and to analyze the changes in a function of two
variables it helps to think of the variables as changing one at a time. That is,
to see how F (x, y) differs from F (a, b), we ﬁrst examine the difference
between F (x, b) and F (a, b) and then look at the difference between
F (x, y) and F (x, b). By temporarily regarding one of the variables as
frozen, we analyze the change in one function of two variables in terms
of the changes in two functions of one variable. In particular, the partial
derivative of F (x, y) with respect to x is found by treating y as a constant
and ﬁnding the derivative of the corresponding function of x. Similarly,
the partial derivative with respect to y is found by treating x as a constant
and ﬁnding the derivative of the corresponding function of y. That’s the
general idea, but there are enough technical details involved that we need
to give more precise statements of what we mean.
Definition 7.1: Given a function F of two variables, suppose that
(a, b) is a point such that F (x, b) is deﬁned for all x in an open interval I
1
containing a, and also that F (a, y) is deﬁned for all y in an open interval
I
2
containing b. We deﬁne
F
1
(a, b) = lim
x→a
F (x, b) −F (a, b)
x −a
and F
2
(a, b) = lim
y→b
F (a, y) −F (a, b)
x −a
provided these limits exist. The quantities F
1
(a, b) and F
2
(a, b) are called
partial derivatives of F; we think of them as functions of the point (a, b).
86 CHAPTER IV THE DERIVATIVE
Just as there are several different notations commonly used for deriva
tives, there are other ways to indicate partial derivatives; the notation we’ve
chosen seems best for our purposes. We’ll often substitute new letters in
place of variables, so we’ve chosen to keep track of the arguments of F
as the ﬁrst or second variable, rather than by the letter used. That is, we
calculate F
1
by differentiating F (x, y) with respect to x, the ﬁrst variable,
and we calculate F
2
by differentiating with respect to the second. Then x
and y can be replaced by numerical expressions as needed.
The condition that allows us to approximate F (x, y) near (a, b) by a
linear function of two variables is again called differentiability. We say
that F is differentiable at (a, b) if its domain includes all points inside some
square
S
δ
= {(x, y) : x −a < δ and y −b < δ}
centered at (a, b) and if there are functions G
1
, G
2
: S
δ
→ R that are
continuous at (a, b) and satisfy
F (x, y) = F (a, b) + (x −a) G
1
(x, y) + (y −a) G
2
(x, y) (4.10)
at all points of S
δ
. (Here we are not assuming that G
1
and G
2
are partial
derivatives of a function G; the subscripts are used to link them to F
1
and
F
2
.) Since (4.10) leads to
and
F (x, b) = F (a, b) + (x −a) G
1
(x, b)
F (a, y) = F (a, b) + (y −b) G
1
(a, y),
differentiability at (a, b) implies the existence of both partial derivatives at
that point, with F
1
(a, b) = G
1
(a, b) and F
2
(a, b) = G
2
(a, b). However,
the existence of F
1
(a, b) and F
2
(a, b) depends only on the values of F
along horizontal and vertical lines through (a, b), while differentiability at
(a, b) involves the values of F (x, y) at all points in S
δ
. But even though
it is possible for both partial derivatives to exist at a point where F is not
differentiable, there is still a way to use partial derivatives to check for
differentiability. The theorem below tells how to do it.
Theorem 7.1: Suppose that the function F has partial derivatives
F
1
and F
2
deﬁned at every point in a square S
δ
centered at (a, b). If F
1
and F
2
are continuous at (a, b), then F is differentiable at (a, b).
Proof : While equation (4.10) doesn’t determine the functions G
1
and G
2
, there’s a simple way to deﬁne them to make it be satisﬁed. Of
course, we want G
1
(a, b) = F
1
(a, b) and G
2
(a, b) = F
2
(a, b), so for
7. IMPLICIT FUNCTIONS AND IMPLICIT DIFFERENTIATION 87
(x, y) = (a, b) we deﬁne
G
1
(x, y) = F
1
(a, b) +
(x −a) E (x, y)
(x −a)
2
+ (y −b)
2
(4.11)
and
G
2
(x, y) = F
2
(a, b) +
(y −b) E (x, y)
(x −a)
2
+ (y −b)
2
, (4.12)
where
E (x, y) = F (x, y) −F (a, b) −(x −a) F
1
(a, b) −(y −b) F
2
(a, b).
Direct substitution shows that (4.10) is satisﬁed, and we may note that this
deﬁnition of G
1
and G
2
makes them continuous at all points other than
(a, b) where F is continuous. To prove continuity at (a, b), we need to
analyze E; that’s where the assumptions about F
1
and F
2
are needed.
For each ﬁxed y ∈ (b −δ, b +δ), we can regard F (x, y) as a differ
entiable function of x on (a −δ, a +δ). So for each x in this interval the
mean value theorem guarantees that there is a number ξ between a and x
with
F (x, y) = F (a, y) + (x −a) F
1
(ξ, y).
Also, we can regardF (a, y) as a differentiable functionof y on (b −δ, b +δ).
So for each y in that interval there is a number η between b and y such that
F (a, y) = F (a, b) + (y −b) F
2
(a, η).
Combining these equations shows
F (x, y) = F (a, b) + (x −a) F
1
(ξ, y) + (y −b) F
2
(a, η),
and therefore E (x, y) is given by
(x −a) [F
1
(ξ, y) −F
1
(a, b)] + (y −b) [F
2
(a, η) −F
2
(a, b)].
Note that if we restrict (x, y) to a smaller square centered at (a, b), then
both (ξ, y) and (a, η) will be in this same square. Hence the continuity of
F
1
and F
2
at (a, b) shows that for any given ε > 0, there is a δ
> 0 such
that the coefﬁcients of (x −a) and (y −b) in the formula for E will both
be in (−ε/2, ε/2) whenever both x −a < δ
and y −b < δ
. Hence for
all such (x, y) we have
E (x, y) <
ε
2
x −a +
ε
2
y −b
< ε
_
(x −a)
2
+ (y −b)
2
. (4.13)
88 CHAPTER IV THE DERIVATIVE
Since both x −a and y −b are bounded by
_
(x −a)
2
+ (y −b)
2
,
using (4.13) in (4.11) and (4.12) shows both
G
1
(x, y) −F
1
(a, b) < ε and G
2
(x, y) −F
2
(a, b) < ε
for all (x, y) with both x −a < δ
and y −b < δ
, proving that G
1
and
G
2
are continuous at (a, b). That completes the proof of the theorem.
Now let’s take another look at the method of implicit differentiation.
We begin by assuming that F is differentiable at (a, b) and that φ(a) = b
with φ differentiable at a. Deﬁne
h(x) = F (x, φ(x)).
We’ll show that h is differentiable at a and calculate h
(a). Since φ
continuous at a, there must be an open interval I = (a −δ, a +δ) on
which all the values of φ(x) are close enough to b to make
h(x) = F (a, b) + (x −a) G
1
(x, φ(x)) + (φ(x) −b) G
2
(x, φ(x)),
with G
1
and G
2
both continuous at (a, b). The continuity of φ at a
guarantees that both G
1
(x, φ(x)) and G
2
(x, φ(x)) deﬁne functions of x
that are continuous at a, and the differentiability of φ at a lets us write
φ(x) −b = (x −a) g (x) for all x ∈ I,
with g continuous at a. So we ﬁnd
h(x) = h(a) + (x −a) G
1
(x, φ(x)) + (x −a) g (x) G
2
(x, φ(x)).
We’ve proved that h is differentiable at a with
h
(a) = G
1
(a, φ(a)) +g (a) G
2
(a, φ(a))
= F
1
(a, b) +φ
(a) F
2
(a, b).
Of course, we can also ﬁnd h
(a) by substituting y = φ(x) in the formula
for F and just using the chain rule along with whatever other differentiation
formulas are appropriate, but it must also be possible to group the results
in the form above.
For implicit differentiation, we also assume that the graph of φ over
I coincides with the graph of the equation F (x, y) = C, so of course
h(x) = C for all x ∈ I and h
(a) = 0. As long as F
2
(a, b) = 0, we can
calculate φ
(a) by simply solving
F
1
(a, b) +φ
(a) F
2
(a, b) = 0.
7. IMPLICIT FUNCTIONS AND IMPLICIT DIFFERENTIATION 89
That’s how implicit differentiation is supposed to work.
Note that we made several signiﬁcant assumptions along the way.
While the differentiability of F at the point (a, b) on the graph of F (x, y) =
C can be veriﬁed by analyzing the function F used to formthe equation, the
assumptions about φ are another matter entirely. Of course, when the graph
of F (x, y) = C is a familiar object and it’s known to be a smooth curve
near (a, b), those assumptions ought to be satisﬁed. But we’d especially
like to use the method when we don’t know much about the graph. That’s
where the next result, commonly known as the implicit function theorem,
comes in. The hypotheses we need include the hypotheses of Theorem 7.1,
not just its conclusion.
Theorem 7.2: Let (a, b) be a point on the graph of F (x, y) = C,
where C is a given constant, and suppose that both partial derivatives F
1
and F
2
of F are deﬁned at all points in a square S centered at (a, b). If F
1
and F
2
are continuous at (a, b) with F
2
(a, b) = 0, then there is a numerical
function φ that is differentiable at a and satisﬁes F (x, φ(x)) = C for all
x in its domain. Also, there is a square S
δ
centered at (a, b) in which the
only points (x, y) on the graph of F (x, y) = C are those on the graph of
φ. The derivative of φ at a is given by
φ
(a) = −F
1
(a, b) /F
2
(a, b).
Proof : Since F satisﬁes the hypotheses of Theorem 7.1, for
(x, y) ∈ S we can write
F (x, y) = C + (x −a) G
1
(x, y) + (y −b) G
2
(x, y),
with G
1
and G
2
continuous at (a, b) and G
2
(a, b) = F
2
(a, b) = 0. By
continuity, we can ﬁnd a smaller square S
δ
centered at (a, b) on which F
2
and G
2
keep the same sign, and the ratio G
1
/G
2
satisﬁes the bound
¸
¸
¸
¸
G
1
(x, y)
G
2
(x, y)
¸
¸
¸
¸
≤ 1 +
¸
¸
¸
¸
F
1
(a, b)
F
2
(a, b)
¸
¸
¸
¸
= M.
Let’s call I the interval (a −δ, a +δ) and J the interval (b −δ, b +δ).
For any c, d ∈ J with c < d and for any ﬁxed x ∈ I, the function
f deﬁned by f (y) = F (x, y) is continuous on [c, d] and has no critical
points in that interval, since f
(y) = F
2
(x, y) = 0. Thus Rolle’s theorem
shows that F (x, c) = F (x, d). So for any x ∈ I, there can never be more
than one y with (x, y) ∈ S
δ
and F (x, y) = C. If we deﬁne A to be the set
of x ∈ I for which there is such a y and call φ(x) that unique value of y,
then φ is a numerical function with domain A, φ(a) = b, and the points
90 CHAPTER IV THE DERIVATIVE
(x, y) in S
δ
on the graph of F (x, y) = C are precisely those on the graph
of φ.
The next step is to prove that A includes an open interval containing
a; that’s an important part of proving that φ is differentiable at a. We pick
r > 0 with b − r and b + r in J, and then note that C must be strictly
between F (a, b −r) and F (a, b +r). Indeed, we have
[F (a, b −r) −C] [F (a, b +r) −C] = −r
2
G
2
(a, b −r) G
2
(a, b +r),
which is negative since G
2
keeps a ﬁxed sign in S
δ
. Since F (x, b −r)
and F (x, b +r) are functions of x that are continuous at a, there must be
an interval (a −ε, a +ε) on which C is always between F (x, b −r) and
F (x, b +r). At each such x, the function f deﬁned by
f (y) = F (x, y)
is continuous on the interval [b −r, b +r] and C is between the values of
f at the endpoints. So the intermediate value theorem guarantees there is a
y in the interval with f (y) = C. But then (x, y) ∈ S
δ
with F (x, y) = C,
and therefore x is in A, the domain of φ.
We complete the proof by showing that φ is differentiable at a. For
each x ∈ A we have
0 = F (x, φ(x)) −F (a, b)
= (x −a) G
1
(x, φ(x)) + (φ(x) −φ(a)) G
2
(x, φ(x)),
and since G
2
(x, φ(x)) = 0 for x ∈ A, we can rewrite this as
φ(x) −φ(a) = −(x −a)
G
1
(x, φ(x))
G
2
(x, φ(x))
.
Our choice of S
δ
guarantees that
φ(x) −φ(a) = x −a
¸
¸
¸
¸
G
1
(x, φ(x))
G
2
(x, φ(x))
¸
¸
¸
¸
≤ M x −a,
so obviously φ is continuous at a. Consequently,
g (x) = −
G
1
(x, φ(x))
G
2
(x, φ(x))
deﬁnes a function on (a −ε, a +ε) that is continuous at a, and for all x in
that interval we have
φ(x) = φ(a) + (x −a) g (x).
EXERCISES 91
Hence φ is differentiable at a, with
φ
(a) = g (a) = −
G
1
(a, φ(a))
G
2
(a, φ(a))
= −
F
1
(a, b)
F
2
(a, b)
.
That completes our proof of the implicit function theorem.
Typically, the hypotheses about F in the implicit function theorem are
satisﬁed at all points in a square centered at (a, b), not just at the single
point, and then we have
φ
(x) = −
F
1
(x, φ(x))
F
2
(x, φ(x))
for all x in an open interval about a. With suitable F
1
and F
2
, we can then
calculate φ
and successively higher derivatives of φ.
Generally we won’t be able to ﬁnd an algebraic formula for the function
φ(x). Nevertheless, we can still form the linearization
φ(a) + (x −a) φ
(a) = b −(x −a)
F
1
(a, b)
F
2
(a, b)
or perhaps form the quadratic approximation
φ(a) + (x −a) φ
(a) +
1
2
(x −a)
2
φ
(a)
to approximate φ(x) near x = a. Such approximations are adequate for
many purposes. Indeed, one of the reasons that calculus is so useful is that
it gives us practical methods for working with unknown functions.
EXERCISES
23. The numerical function F deﬁned by
F (x, y) =
_
¸
_
¸
_
xy
x
2
+y
2
, (x, y) = (0, 0)
0, (x, y) = (0, 0)
satisﬁes F
1
(0, 0) = F
2
(0, 0) = 0 since F vanishes along both coor
dinate axes. Show that F is neither continuous nor differentiable at
(0, 0).
24. For (a, b) any point except the origin, prove that the equation
x
2
y +y
3
= b
_
a
2
+b
2
_
deﬁnes y as a differentiable function φ(x) on some open interval
about a. Find the linearization of φ(x) near x = a.
V
THE RIEMANN INTEGRAL
O
ne of the early triumphs of calculus was a general method for
ﬁnding areas of plane regions bounded by curves. When f is
a positive, continuous function on [a, b], the area of the region
bounded by the graph of f, the xaxis, and the vertical lines x = a
and x = b is usually called the area below the graph of f. This area
can be approximated by slicing the region into strips with vertical lines,
estimating the areas of the strips by areas of rectangles, and then adding
the areas of the strips. By reducing the width of the strips, the error in
such an approximation can be made arbitrarily small, and the exact area
is the limiting value of such sums as the widths of all the strips involved
approach zero. The methods of calculus provide systematic ways to work
with such limits and often let us determine their exact values.
This approach to the problemof ﬁnding areas is an example of integra
tion, a powerful technique used in all sorts of measurements in science and
mathematics. It’s applicable to just about every measurement that theoret
ically can be performed by decomposing something into parts, measuring
92
1. AREAS AND RIEMANN SUMS 93
the parts, and adding the results. For example, volume, mass, and energy
are such measurements, but pressure and temperature are not. As we look
for ways to compute such measurements, the way we ﬁnd areas serves as
a useful model, and helps us understand the properties of the operations
involved.
1 AREAS AND RIEMANN SUMS
We’ll study the theory of integration in the form developed by the German
mathematician Bernhard Riemann in the nineteenth century. While the
basic ideas had been around for some time, he is generally credited with
being the ﬁrst to formulate them in a way that could be applied to functions
with discontinuities and still be rigorously correct. Later mathematicians
found a need to extend the notion of integration to situations falling outside
Riemann’s theory, and several newer theories of integration are commonly
encountered in higher mathematics; some of these theories are still under
development.
Our ﬁrst task is to deﬁne the terms we’ll work with in this chapter. We’ll
assume that a < b, and that we have a numerical function f whose domain
includes [a, b]. Apartition of [a, b] is a splitting of it into a ﬁnite number of
nonoverlapping closed subintervals. The word nonoverlapping is carefully
chosen; intervals that share interior points overlap, but closed intervals that
share only an endpoint do not. Partitions of [a, b] are conveniently speciﬁed
in terms of a ﬁnite sequence {x
k
}
n
k=0
with
a = x
0
< x
1
< x
2
< · · · < x
n
= b.
Then the partition consists of n subintervals, conveniently numbered left
to right as
I
1
= [x
0
, x
1
], I
2
= [x
1
, x
2
], . . . , I
n
= [x
n−1
, x
n
].
We use the symbol ∆x
k
= x
k
−x
k−1
for the length of the kth subinterval
I
k
. The mesh of the partition is our name for the largest of the lengths
{∆x
1
, ∆x
2
, ..., ∆x
n
}.
We call points {x
∗
1
, x
∗
2
, . . . , x
∗
n
} sampling points for the partition provided
that x
∗
k
∈ I
k
for each k = 1, 2, . . . , n. We call a sum
n
k=1
f (x
∗
k
) ∆x
k
= f (x
∗
1
) ∆x
1
+f (x
∗
2
) ∆x
2
+· · · +f (x
∗
n
) ∆x
n
94 CHAPTER V THE RIEMANN INTEGRAL
a Riemann sum for f associated with the partition {x
k
}
n
k=0
as long as the
points {x
∗
k
}
n
k=1
are sampling points for the partition. In general, the value
of a Riemann sum depends on the partition and sampling points used, as
well as on the function f and the interval [a, b].
When f is a positive, continuous function on [a, b] and we’re trying to
ﬁnd the area belowits graph, the Riemann sumformed above represents the
sum of the areas of rectangles R
1
, R
2
, . . . , R
n
, with R
k
having I
k
for its
base and (x
∗
k
, f (x
∗
k
)) a point in its top. Sketches suggest that the Riemann
sums approach the area below the graph as we consider partitions with
smaller meshes. But when f is neither positive nor continuous, the picture
is much more complicated, and the Riemann sums for f need not approach
a single number. To deal with this difﬁculty, we adopt a common strategy
in mathematics—we use the operation we hope to perform to deﬁne a class
of functions and then look for ways to improve our understanding of this
class.
Definition 1.1: For f a numerical function whose domain in
cludes the nondegenerate closed interval [a, b], we say that f is Riemann
integrable over [a, b] if there is a number, called the integral of f over
[a, b] and denoted by
_
b
a
f (x) dx, such that for each given ε > 0, there
is a δ > 0 with the property that whenever {x
k
}
n
k=0
is a partition of [a, b]
having mesh less than δ, every Riemann sum
n
k=1
f (x
∗
k
) ∆x
k
associated
with it satisﬁes
¸
¸
¸
¸
¸
n
k=1
f (x
∗
k
) ∆x
k
−
_
b
a
f (x) dx
¸
¸
¸
¸
¸
< ε.
The process that determines the value of
_
b
a
f (x) dx is called integration.
That’s the traditional deﬁnition of the integral. Not long ago it was
given in roughly that form in just about every calculus book. It’s an
incredibly complicated deﬁnition, and many newer books are attempting to
replace it by something simpler. Its virtue is that when we know a function
to be Riemann integrable, we have readymade schemes for approximating
its integral and a powerful tool for proving things about it. But it can be
extremely difﬁcult to show that a given function satisﬁes the deﬁnition.
However, it’s easy to verify directly that constant functions are inte
grable over every closed interval, with
_
b
a
C dx = C (b −a) for any constant C.
1. AREAS AND RIEMANN SUMS 95
Indeed, when f (x) = C we note that C (b −a) is the value of every
Riemann sum associated with every partition of [a, b] because
n
k=1
f (x
∗
k
) ∆x
k
=
n
k=1
C∆x
k
= C
_
n
k=1
∆x
k
_
= C (b −a).
Usually we’re not so lucky, and the values of the Riemann sums are much
harder to ﬁnd.
The deﬁnition is better suited to proving things about integrable func
tions. For example, it’s easy to prove that every linear combination of
integrable functions is integrable. More precisely, if f and g are Riemann
integrable over [a, b] and h = αf + βg with α and β constants, then h is
Riemann integrable over [a, b] and
_
b
a
[αf (x) +βg (x)] dx = α
_
b
a
f (x) dx +β
_
b
a
g (x) dx.
We prove it by showing that α
_
b
a
f (x) dx + β
_
b
a
g (x) dx satisﬁes the
deﬁnition of the Riemann integral of h over [a, b]. We ﬁrst note that for an
arbitrary partition of [a, b] and any choice of the sampling points, we can
rewrite the Riemann sum for h in terms of the Riemann sums for f and g:
n
k=1
h(x
∗
k
) ∆x
k
= α
_
n
k=1
f (x
∗
k
) ∆x
k
_
+β
_
n
k=1
g (x
∗
k
) ∆x
k
_
.
Given ε > 0, our rules for approximate arithmetic tell us that there is an
ε
> 0 such that
¸
¸
¸
¸
αu +βv −
_
α
_
b
a
f (x) dx +β
_
b
a
g (x) dx
_¸
¸
¸
¸
< ε
for all u within ε
of
_
b
a
f (x) dx and for all v within ε
of
_
b
a
g (x) dx. For
both f and g Riemann integrable, there must be a δ > 0 such that
u =
n
k=1
f (x
∗
k
) ∆x
k
and v =
n
k=1
g (x
∗
k
) ∆x
k
will satisfy these conditions whenever the mesh of P is smaller than δ.
This will guarantee that
¸
¸
¸
¸
¸
n
k=1
h(x
∗
k
) ∆x
k
−
_
α
_
b
a
f (x) dx +β
_
b
a
g (x) dx
_
¸
¸
¸
¸
¸
< ε.
96 CHAPTER V THE RIEMANN INTEGRAL
Here’s another simple consequence of the deﬁnition: if f and g are
Riemann integrable over [a, b] with f (x) ≤ g (x) for all x ∈ [a, b], then
_
b
a
f (x) dx ≤
_
b
a
g (x) dx.
To prove it, we ﬁrst note that if f (x) ≤ g (x) for all x ∈ [a, b], then for any
choice of sampling points associated with any partition of [a, b] we have
n
k=1
f (x
∗
k
) ∆x
k
≤
n
k=1
g (x
∗
k
) ∆x
k
.
If both f and g are Riemann integrable, then for any ε > 0 there is a
partition with mesh small enough to guarantee that the Riemann sums are
within ε of the corresponding integrals. Consequently,
_
b
a
f (x) dx −ε <
_
b
a
g (x) dx +ε
for every ε > 0, and that forces the desired inequality to be true.
The last two arguments give a hint as to why such a complicated deﬁ
nition developed. When we can approximate
_
b
a
f (x) dx and
_
b
a
g (x) dx
separately by Riemann sums, we want to be able to approximate them
simultaneously—that is, by using Riemann sums corresponding to the
same partition with the same sampling points. Part of the price we pay for
that convenience is that to prove integrability, we must establish a bound
that is valid for every choice of sampling points with every partition of
sufﬁciently small mesh, not just for one convenient choice. But that’s more
of a nuisance than an overwhelming obstacle. However, showing that the
deﬁnition is satisﬁed requires a value for the integral, and the value can be
hard to ﬁnd. It wasn’t at all hard in the cases we just considered; that’s why
they were chosen at this stage. But when rounding up the usual suspects
fails to produce a likely candidate for the value of the integral, some real
detective work can be required.
When f is the derivative of a function F that satisﬁes the hypotheses
of the mean value theorem on [a, b], then F (b) − F (a) is the only real
possibility for
_
b
a
f (x) dx. With any partition, each sampling point can
presumably be chosen to satisfy the conclusion of the mean value theorem
on the corresponding subinterval; that is
F (x
k
) −F (x
k−1
) = f (x
∗
k
) (x
k
−x
k−1
) = f (x
∗
k
) ∆x
k
.
EXERCISES 97
The corresponding Riemann sum is then exactly F (b) −F (a). But there’s
much more to the story of the Riemann integral because many Riemann in
tegrable functions are not derivatives and some derivatives are not Riemann
integrable.
Many calculus texts refer to deﬁnite integrals and indeﬁnite integrals.
Our use of integral corresponds to the deﬁnite integral; we prefer to use an
tiderivative instead of indeﬁnite integral. That’s mainly to emphasize that
the concept of integration we’ve deﬁned is unrelated to the manipulations
used to discover antiderivatives.
EXERCISES
1. Use the intermediate value theorem to prove that if f is positive
and continuous on [a, b], then there is an x
∗
∈ [a, b] such that the
area below the graph of f is exactly f (x
∗
) (b −a). Why don’t we
calculate this area by ﬁrst ﬁnding x
∗
?
2. Prove that if f is positive and continuous on [a, b], then for every
partition {x
k
}
n
k=0
of [a, b] there are sampling points {x
∗
k
}
n
k=1
such
that the corresponding Riemann sumis exactly equal to the area below
the graph of f.
3. For {x
k
}
n
k=0
an arbitrary partition of [a, b] and ¯ x
k
the midpoint of
I
k
, show that
n
k=1
¯ x
k
∆x
k
=
1
2
_
b
2
−a
2
_
. Then use the deﬁnition
to prove that x is Riemann integrable over [a, b] with
_
b
a
xdx =
1
2
_
b
2
−a
2
_
. The key is to observe that
n
k=1
x
∗
k
∆x
k
−
1
2
_
b
2
−a
2
_
=
n
k=1
(x
∗
k
− ¯ x
k
) ∆x
k
.
Then x
∗
k
− ¯ x
k
 is easy to bound in terms of the mesh of the partition.
4. Suppose that a < b < c, and that f is known to be Riemann integrable
over [a, b], [b, c], and [a, c]. Use the deﬁnition of the integral to prove
that
_
b
a
f (x) dx +
_
c
b
f (x) dx =
_
c
a
f (x) dx.
(It isn’t really necessary to assume Riemann integrability over all
three intervals, but without that assumption the argument is much
more difﬁcult.)
98 CHAPTER V THE RIEMANN INTEGRAL
2 SIMPLIFYING THE CONDITIONS FOR INTEGRABILITY
At the end of the previous section we commented on the difﬁculties in
showing that a given function satisﬁes the deﬁnition of Riemann integra
bility. Here we’ll see that some of these difﬁculties can be bypassed. In
particular, we’ll develop a new condition that doesn’t involve the value of
the integral.
We begin by considering the set of all Riemann sums for f associated
with a given partition of [a, b]. To make it easier to think of the partition as
a variable, let’s call
P = {x
k
}
n
k=0
with a = x
0
< x
1
< x
2
< · · · < x
n
= b
and then call R(P) the set of all Riemann sums for f associated with P
that we can form by choosing sampling points; that is,
R(P) =
_
n
k=1
f (x
∗
k
) ∆x
k
: x
∗
1
∈ I
1
, . . . , x
∗
n
∈ I
n
_
.
Our ﬁrst observation is that every function that is Riemann integrable
over [a, b] must be bounded on [a, b]. In fact, if we’ve got a partition P
such that every element of R(P) is within ε of some number, that alone
forces f to be bounded on [a, b]. Since different choices of the sampling
point x
∗
k
∈ I
k
can’t change the resulting Riemann sum by more that 2ε, it
follows that
diam{f (x) : x ∈ I
k
} ≤ 2ε/∆x
k
for k = 1, . . . , n.
Thus f is bounded on each I
k
and [a, b] is formed from only ﬁnitely many
of the subintervals I
k
, so we see that f must be bounded on all of [a, b];
that is, we can assume that there are numbers m and M with
m ≤ f (x) ≤ M for all x ∈ [a, b].
Conversely, if no such numbers exist, then we know that f cannot be
Riemann integrable over [a, b].
Our next observation is that we don’t really need to consider arbitrary
partitions with mesh less than some number; a simpler version of the deﬁ
nition of Riemann integrability can be given that only involves considering
one partition at a time. The theorem below makes this precise.
Theorem 2.1: Suppose that f is a numerical function deﬁned on
[a, b], and that S is a number with the property that for each ε > 0, there
2. SIMPLIFYING THE CONDITIONS FOR INTEGRABILITY 99
is a partition P of [a, b] with R(P) ⊂ (S −ε, S +ε). Then f is Riemann
integrable over [a, b] and S =
_
b
a
f (x) dx.
It’s obvious that the hypotheses of this theorem are satisﬁed whenever
f satisﬁes our original deﬁnition of Riemann integrability, so this theorem
can serve as an alternate deﬁnition once we’ve proved it. Proving it depends
on a careful analysis of how R(P) can change as we change P.
The simplest way to change a partition P
0
is to reduce it to a new
partition P by deleting one of the intermediate points x
k
, so that two
adjacent subintervals are collapsed to a single subinterval. In the process,
we replace the two terms f (x
∗
k
) ∆x
k
and f
_
x
∗
k+1
_
∆x
k+1
by a single
term corresponding to a new subinterval of length ∆x
k
+ ∆x
k+1
. Since
x
∗
k
and x
∗
k+1
are possible sampling points for the new subinterval, and the
sum of the replaced terms is always between f (x
∗
k
) (∆x
k
+ ∆x
k+1
) and
f
_
x
∗
k+1
_
(∆x
k
+ ∆x
k+1
), we see that each Riemann sum in R(P
0
) lies
between two Riemann sums in R(P). Consequently, R(P
0
) is a subset
of each interval that includes R(P), and that remains true if we further
change P by deleting additional points.
The changes to P that we really want to consider are not reductions;
we’re more interested in reversing the process by adding new intermediate
points. A partition formed from P by inserting additional intermediate
points is called a reﬁnement of P. Since we can always reduce a reﬁnement
to get back the original partition, the previous paragraph tells us howthe set
of Riemann sums can change when we reﬁne the partition. In particular,
whenever P
is a reﬁnement of P, we see that R(P
) must be a subset of
any interval that contains R(P).
Proof : Now we prove the theorem itself. We assume that f
satisﬁes the stated hypotheses, then prove that f satisﬁes the deﬁnition of
Riemann integrability with S =
_
b
a
f (x) dx. That is, we show that for any
given ε > 0, there is a δ > 0 such that every partition P with mesh less
than δ must satisfy
R(P) ⊂ (S −ε, S +ε).
Our hypotheses tell us there must be a partition P
0
with
R(P
0
) ⊂
_
S −
1
2
ε, S +
1
2
ε
_
;
for the moment, consider P and δ as known. When we reﬁne P by inserting
all the points of P
0
that aren’t already there, the resulting partition P
is
100 CHAPTER V THE RIEMANN INTEGRAL
also a reﬁnement of P
0
, so
R
_
P
_
⊂
_
S −
1
2
ε, S +
1
2
ε
_
.
The idea is to make sure that every Riemann sum in R(P) is within
1
2
ε of
one in R(P
). When we add a single point to P, we need to choose only
one newsampling point to produce a Riemann sumfor the reﬁned partition.
So if all the values of f are in [m, M], we can get a new Riemann sum that
is within (M −m) δ of the old one. Extending this analysis, we see that
if n is the number of points used to deﬁne P
0
, then we add no more than n
new points to P to produce P
. So every Riemann sum in R(P) must be
within n(M −m) δ of a Riemann sum in R(P
). Choosing δ > 0 with
n(M −m) δ ≤
1
2
ε
completes the proof.
Our next task is to develop another characterization of the value of the
Riemann integral. The example of ﬁnding the area below the graph of a
positive continuous function is a useful model. Assuming only that f is a
bounded function deﬁned on [a, b], for any partition P we may deﬁne
m
k
= inf {f (x) : x ∈ I
k
} and M
k
= sup{f (x) : x ∈ I
k
}
and use them to form
L(P) =
n
k=1
m
k
∆x
k
and U (P) =
n
k=1
M
k
∆x
k
,
commonly called the lower sum for f relative to P and the upper sum for
f relative to P.
In the case of a positive continuous function f, L(P) represents the
area of an inscribed region made up of rectangles and U (P) represents
the area of a circumscribed region, so the area under the graph must be
somewhere between L(P) and U (P). In the general case, we must still
have
R(P) ⊂ [L(P), U (P)]
because Theorem 3.4 in Chapter 1 tells us that
L(P) = inf R(P) and U (P) = supR(P).
2. SIMPLIFYING THE CONDITIONS FOR INTEGRABILITY 101
We’ll see that when f is Riemann integrable,
_
b
a
f (x) dx can be deﬁned
as the one number common to all the intervals of the form [L(P), U (P)].
Of course, that involves proving that there is such a number.
It’s not at all obvious why there should be any number common to all
the intervals in the family
I ={[L(P), U (P)] : P a partition of [a, b]},
but Theorem 3.2 in Chapter 1 comes to our rescue. We just need to show
that every two intervals in I have points in common. In fact, the intersection
of any two intervals in I must always contain a third interval in I. Two
intervals I
1
and I
2
in I must correspond to partitions P
1
and P
2
of [a, b],
and then for P any reﬁnement of both P
1
and P
2
we know that R(P) is a
subset of both I
1
and I
2
. Consequently,
I = [L(P), U (P)] ∈ I with I ⊂ I
1
∩ I
2
.
When I includes arbitrarily short intervals, there must be exactly one
point common to all the intervals in I. That’s the key to our ﬁnal charac
terization of Riemann integrability given in the theorem below.
Theorem 2.2: Let f be a bounded numerical function on [a, b].
Then f is Riemann integrable over [a, b] if for each ε > 0, there is a
partition P of [a, b] such that
diamR(P) = U (P) −L(P) < ε.
In that case,
_
b
a
f (x) dx is the one number common to all the intervals in
the family
I = {[L(P), U (P)] : P a partition of [a, b]}.
Proof : For S the one number common to all the intervals in the
family I, when U (P) −L(P) < ε we have
R(P) ⊂ [L(P), U (P)] ⊂ (S −ε, S +ε).
So the conclusion follows from Theorem 2.1. It’s worth noting that ev
ery Riemann integrable function must satisfy the hypotheses of Theorem
2.2.
This characterization of the value of the integral makes it obvious why
m(b −a) ≤
_
b
a
f (x) dx ≤ M (b −a)
102 CHAPTER V THE RIEMANN INTEGRAL
when f is Riemann integrable with m ≤ f (x) ≤ M for all x ∈ [a, b]; just
look at L(P) and U (P) for P ={a, b}.
EXERCISES
5. For a given bounded function f on [a, b] and partition P of [a, b],
show how to express the diameter of R(P) in terms of the quantities
D
k
= diam{f (x) : x ∈ I
k
} = sup {f (x) −f (y) : x, y ∈ I
k
}.
6. Prove that if f is Riemann integrable over [a, b], then so is f. (The
preceding exercise is useful.) When is
_
b
a
f (x) dx =
¸
¸
¸
¸
_
b
a
f (x) dx
¸
¸
¸
¸
?
7. Prove that if f is Riemann integrable over [a, b], then so are f
+
and
f
−
, where f = f
+
−f
−
with
f
+
(x) =
_
f (x) , f (x) ≥ 0
0, f (x) < 0
and
f
−
(x) =
_
0, f (x) ≥ 0
−f (x) , f (x) < 0.
8. Prove that if f and g are Riemann integrable over [a, b], then so is fg.
The simplest way is to write f (x) g (x) −f (y) g (y) as
f (x) [g (x) −g (y)] + [f (x) −f (y)] g (y)
to ﬁnd a useful bound for
diam{f (x) g (x) : x ∈ [x
k−1
, x
k
]}.
3 RECOGNIZING INTEGRABILITY
Now that we’ve developed Theorem 2.2, we’ll use it in a way that will
just about make it obsolete. We’ll show that large classes of functions
are Riemann integrable. Then we can establish the integrability of most
of the integrable functions we encounter by inspection; we need only
recognize their membership in such a class. For example, we can often
recognize a function as increasing or decreasing over the entire interval of
3. RECOGNIZING INTEGRABILITY 103
integration; such functions are called monotonic. We can also recognize
many functions as continuous. The next two theorems state that every
function in either of these classes is Riemann integrable.
Theorem 3.1: Every monotonic function deﬁned on [a, b] is Rie
mann integrable over [a, b].
Proof : We’ll use Theorem 2.2. When f is monotonic, all the
values of f on [a, b] will be between f (a) and f (b), so we know that f is
bounded on [a, b]. If we restrict our attention to partitions involving only
subintervals of equal lengths, we discover that U (P) −L(P) is quite easy
to calculate.
Let’s begin by assuming that f is increasing on [a, b] and form a
partition P = {x
k
}
n
k=0
by deﬁning
x
k
= a +
k
n
(b −a) for k = 0, 1, . . . , n.
Since f is increasing, we recognize that
m
k
= inf {f (x) : x ∈ I
k
} = f (x
k−1
)
and
M
k
= sup{f (x) : x ∈ I
k
} = f (x
k
).
By design,
∆x
k
=
b −a
n
for all k.
We may now compute
L(P) =
n
k=1
m
k
∆x
k
= [f (x
0
) +f (x
1
) +· · · +f (x
n−1
)] ·
b −a
n
and
U (P) =
n
k=1
M
k
∆x
k
= [f (x
1
) +f (x
2
) +· · · +f (x
n
)] ·
b −a
n
.
Subtracting, we ﬁnd that the diameter of R(P) is
U (P) −L(P) = [f (x
n
) −f (x
0
)] ·
b −a
n
=
1
n
[f (b) −f (a)] (b −a) .
104 CHAPTER V THE RIEMANN INTEGRAL
Clearly, we can make this quantity as small as we please by simply choosing
n large enough, so we’ve satisﬁed our condition for integrability.
The argument for decreasing functions is only slightly different; m
k
be
comes f (x
k
) and M
k
becomes f (x
k−1
), so the formula for U (P)−L(P)
becomes
1
n
[f (a) −f (b)] (b −a).
Theorem 3.2: Every continuous function on [a, b] is Riemann inte
grable over [a, b].
Proof : The integrability of every function that is continuous on
[a, b] is a much deeper result, but we did the hard part in Chapter 2 when
we proved Theorem 5.2. Given ε > 0, whenever f is continuous on [a, b]
we can ﬁnd a partition P such that M
k
−m
k
≤ ε for all k. We just choose
δ such that every x, y ∈ [a, b] with x −y < δ satisfy f (x) −f (y) < ε,
and then choose any partition P with mesh less than δ. For such a partition
we have
U (P) −L(P) =
n
k=1
(M
k
−m
k
) ∆x
k
≤ ε
n
k=1
∆x
k
= ε (b −a).
Since we can make this last quantity arbitrarily small by choosing ε, we’ve
satisﬁed the condition for integrability.
Here’s another theorem that’s easy to prove, thanks to our simpliﬁed
condition for integrability. It’s used primarily for proving things about
general integrable functions; with speciﬁc functions we can usually rec
ognize integrability over the subinterval in the same way we recognized
integrability over the larger interval.
Theorem 3.3: Every function that is Riemann integrable over a
closed interval [a, b] is Riemann integrable over every closed subinterval
of [a, b].
Proof : Suppose that f is Riemann integrable over [a, b] and that
[c, d] is a given subinterval of [a, b]. Since f must be bounded on [a, b],
of course it also bounded on [c, d]. Given ε > 0, we need to show there
is a partition P of [c, d] with U (P) − L(P) < ε. We know that there
is a partition P
of [a, b] with U (P
) − L(P
) < ε and that P
has a
reﬁnement P
= {x
k
}
n
k=0
that includes c and d. We form P from P
by
deleting all the points before c or after d. Since all the terms in the sum
n
k=1
[M
k
−m
k
] ∆x
k
are nonnegative,
U (P) −L(P) ≤ U
_
P
_
−L
_
P
_
≤ U
_
P
_
−L
_
P
_
< ε.
3. RECOGNIZING INTEGRABILITY 105
That proves the theorem.
The theorem has a useful companion, often used to evaluate integrals
when the formula for f (x) changes at some point in the interval of inte
gration. It also allows us to recognize the integrability of functions that
are neither monotonic nor continuous over the entire interval; we split the
interval into subintervals where one or the other of those conditions is
satisﬁed.
Theorem 3.4: Suppose that a < b < c, and that f is Riemann
integrable over both [a, b] and [b, c]. Then f is Riemann integrable over
[a, c] as well, and
_
c
a
f (x) dx =
_
b
a
f (x) dx +
_
c
b
f (x) dx.
Proof : Given ε > 0, we can ﬁnd a partition P
1
of [a, b] with
R(P
1
) ⊂
__
b
a
f (x) dx −
1
2
ε,
_
b
a
f (x) dx +
1
2
ε
_
as well as a partition P
2
of [b, c] with
R(P
2
) ⊂
__
c
b
f (x) dx −
1
2
ε,
_
c
b
f (x) dx +
1
2
ε
_
.
We combine P
1
and P
2
to form a partition P of [a, c], and then observe
that every Riemann sum in R(P) can be split into one Riemann sum in
R(P
1
) plus another in R(P
2
), so that R(P) ⊂ (S −ε, S +ε) with
S =
_
b
a
f (x) dx +
_
c
b
f (x) dx.
Hence the desired conclusion follows from Theorem 2.1.
Our ﬁnal theorem in this section is useful for integrating discontinu
ous functions. When combined with the previous theorem, it shows that
changing the value of a function at a ﬁnite number of points in an interval
affects neither the integrability nor the integral of the function.
Theorem 3.5: Suppose that f is Riemann integrable over [a, b], and
that g is a function deﬁned on [a, b] that agrees with f on (a, b). Then g is
Riemann integrable over [a, b], and
_
b
a
g (x) dx =
_
b
a
f (x) dx.
106 CHAPTER V THE RIEMANN INTEGRAL
Proof : Once again we use Theorem 2.1. We know that f must be
bounded on [a, b] and then g must be bounded as well. When m ≤ f (x) ≤
M for all x ∈ [a, b], we see that g is bounded by
m
= min {m, g (a), g (b)} and M
= max {M, g (a), g (b)}.
Changing f to g in any Riemann sum associated with a partition P can’t
change the sum by more than 2 (M
−m
) times the mesh of the partition,
and that’s the key to the proof. When ε > 0 is given, we ﬁnd δ with
2 (M
−m
) δ ≤
1
2
ε. Then we choose a partition P with mesh at most
δ and having the property that every Riemann sum for f associated with
P is within
1
2
ε of
_
b
a
f (x) dx. Every Riemann sum for g associated with
P will then be within ε of
_
b
a
f (x) dx, and that’s all we needed to show.
Because of the last theorem, it’s common to consider a function as
Riemann integrable over an interval in some cases where a ﬁnite number
of points in the interval are missing from the domain of the function. For
example, if f (x) = x then f
(x) isn’t deﬁned at the origin, yet we can
consider f
to be integrable over [−1, 1] because there is an integrable
function that agrees with it everywhere except at a single point. But the
theorem is no panacea. For example, 1/x is continuous everywhere except
at the origin, but it isn’t integrable over [0, 1]; the problemis not the absence
of a value at x = 0 but the behavior of the function near that point.
EXERCISES
9. Use the intermediate value theorem to show that if f is continuous on
[a, b], then there is at least one c ∈ (a, b) such that
_
b
a
f (x) dx = f (c) (b −a).
This is the simplest case of the mean value theorem for integrals.
10. A more general version of the mean value theorem for integrals states
that if f is continuous on [a, b] and if g is positive and Riemann
integrable over [a, b], then there is at least one c ∈ (a, b) such that
_
b
a
f (x) g (x) dx = f (c)
_
b
a
g (x) dx.
Of course, the theorem must also apply when g is negative on [a, b]
because −g would then be positive. Show by example why it is
necessary to assume that g keeps the same sign throughout [a, b].
4. FUNCTIONS DEFINED BY INTEGRALS 107
11. We say that a function f deﬁned on [a, b] is a step function if there is a
partition {x
k
}
n
k=1
of [a, b] such that f is constant on each subinterval
(x
k−1
, x
k
), k = 1, . . . , n. Explain why step functions are integrable,
and ﬁnd a formula for
_
b
a
f (x) dx when f is a step function.
12. Prove that every numerical function that is continuous on (a, b) and
bounded on [a, b] is Riemann integrable over [a, b]. (Most of the
terms in the sum for U (P) − L(P) correspond to a partition of a
closed subinterval of (a, b). Use that observation to prove that P can
be chosen to make U (P) −L(P) arbitrarily small.)
13. Generalize the previous exercise to prove that a bounded function f
deﬁned on [a, b] is integrable if it has only ﬁnitely many discontinu
ities.
4 FUNCTIONS DEFINED BY INTEGRALS
So far we’ve generally been assuming that our interval of integration was
ﬁxed, although we have considered the possibility of integrating over subin
tervals. Now we wish to study new functions whose values are given by
integrating a ﬁxed function over a variable subinterval of a ﬁxed interval
I. We assume that we have a numerical function f whose domain includes
all of I, and that f is Riemann integrable over [a, b] whenever [a, b] ⊂ I.
When I is a closed interval we need f to be Riemann integrable over I,
but when I isn’t closed we need not assume that f is even bounded on
I. For example, when f is either continuous or monotonic on I it will be
Riemann integrable over every closed subinterval of I. We won’t worry
further about how we might be able to tell if this condition is satisﬁed;
we’ll simply state it as a hypothesis.
The ﬁrst order of business is to extend the deﬁnition of
_
t
a
f (x) dx
to every pair of numbers a, t ∈ I, not just those with a < t. Following
standard practice, we agree that
_
t
a
f (x) dx = −
_
a
t
f (x) dx when t < a
and
_
a
a
f (x) dx = 0.
With that understanding, for all a, u, v ∈ I we have
_
v
u
f (x) dx =
_
v
a
f (x) dx −
_
u
a
f (x) dx (5.1)
108 CHAPTER V THE RIEMANN INTEGRAL
regardless of their position in the interval. Proving this relationship is
simple but tedious; after we deal with the trivial cases where some of the
endpoints are equal there are still six cases to consider. When a < u < v,
equation (5.1) is a statement about integrals over [a, u], [u, v, ], and [a, v],
and it agrees with the formula proved in Theorem 3.4. The other cases
require an additional step. For example, when u < v < a we need to
ﬁrst rewrite equation (5.1) in terms of integrals over [u, v], [v, a], and [u, a]
before appealing to Theorem 3.4.
In view of (5.1), every function deﬁned by integrating f over a variable
interval can be treated in terms of
_
t
a
f (x) dx with a single ﬁxed value of
a. We simply use our rules for working with combinations of functions.
So all the results we’ll ever need can be stated conveniently in the theorem
below.
Theorem 4.1: Let I be an interval and let f be a numerical function
whose domain includes I. Suppose that f is Riemann integrable over every
closed subinterval [a, b] of I. Then the formula
F (t) =
_
t
a
f (x) dx, t ∈ I
deﬁnes a continuous numerical function on I. Moreover, F is differentiable
at each point in the interior of I where f is continuous, and at those points
it satisﬁes
F
(t) = f (t).
Before we prove this theorem, let’s try to understand its signiﬁcance.
It’s almost as important as the formula
f
(a) = lim
x→a
f (x) −f (a)
x −a
.
In particular, when f is continuous on I, it shows us how we can deﬁne
an antiderivative of f on I, that is, a continuous function on I whose
derivative is f. It’s often useful to be able to refer to a formula for the
solution to a problem, and Theorem 4.1 gives the formula used for the
problem of ﬁnding antiderivatives. It’s also a formula that can be used
for practical calculations. Just as we can approximate the value of the
derivative of f at an arbitrary point a by evaluating a difference quotient
[f (x) −f (a)] / (x −a) with x close to a but not exactly equal to it, we can
approximate a value of a particular antiderivative by choosing a partition
and evaluating a Riemann sum.
4. FUNCTIONS DEFINED BY INTEGRALS 109
Proof : We begin the proof by checking the continuity of F at an
arbitrary point b in I. The key is to note that for all t ∈ I,
F (t) = F (b) +
_
t
b
f (x) dx.
For any b ∈ I we can ﬁnd an r > 0 such that I
b
= [b −r, b +r] ∩ I is
a closed subinterval of I; it doesn’t matter whether b is an interior point
or an endpoint of the interval I. Our hypotheses say that f is Riemann
integrable over I
b
, and so there is a positive number M such that
−M ≤ f (x) ≤ M for all x ∈ I
b
.
That gives us a simple way to bound the integral of f over subintervals of
I
b
, and for all t ∈ I
b
we have
F (t) −F (b) =
¸
¸
¸
¸
_
t
b
f (x) dx
¸
¸
¸
¸
≤ M t −b.
To satisfy the deﬁnition of continuity of F at b, when we’re given ε > 0
we choose δ to be the smaller of r and ε/M.
Next we assume that b is in the interior of I and that f is continuous
at b. Then we prove that F is differentiable at b with F
(b) = f (b). This
time we can choose r > 0 such that (b −r, b +r) ⊂ I, and for all t in that
interval except b itself we can write
F (t) = F (b) +
_
t
b
f (x) dx
= F (b) + (t −b)
_
1
t −b
_
t
b
f (x) dx
_
.
So we need to show that the expression in brackets deﬁnes a function with
a removable discontinuity at t = b and that f (b) is the value we use to
remove the discontinuity. We ﬁrst write
_
t
b
f (x) dx =
_
t
b
f (b) dx +
_
t
b
[f (x) −f (b)] dx
= (t −b) f (b) +
_
t
b
[f (x) −f (b)] dx.
Since f is continuous at b, for any given ε > 0 there is a δ > 0 such that
−
1
2
ε < f (x) −f (b) <
1
2
ε for every x ∈ (b −δ, b +δ) ∩ I,
110 CHAPTER V THE RIEMANN INTEGRAL
and it doesn’t hurt to take δ ≤ r. Then for t = b in (b −δ, b +δ), we have
¸
¸
¸
¸
1
t −b
_
t
b
f (x) dx −f (b)
¸
¸
¸
¸
=
¸
¸
¸
¸
1
t −b
_
t
b
[f (x) −f (b)] dx
¸
¸
¸
¸
≤
1
2
ε < ε.
That’s just what we need to complete the proof.
While Theorem 4.1 guarantees the differentiability of
_
t
a
f (x) dx at
points where f is continuous, it says nothing about differentiability at other
points. Without continuity, differentiability may well fail; for example,
consider
f (x) =
_
1, x ≥ 0
−1, x < 0.
In this case we easily ﬁnd
_
t
0
f (x) dx = t,
and t is not differentiable at 0, the discontinuity of f. On the other hand,
if we deﬁne
g (x) =
_
1, x = a
2, x = a,
then we ﬁnd
d
dt
_
t
0
g (x) dx =
d
dt
(t) = 1 for all t ∈ R.
So in this case, differentiating the function deﬁned by the integral removes
the discontinuity in the function integrated.
The chain rule provides a useful extension of Theorem 4.1. If u(t)
and v (t) are points in the interior of I at which f is continuous, and u and
v are differentiable functions, then
d
dt
_
v(t)
u(t)
f (x) dx = f (v (t)) v
(t) −f (u(t)) u
(t).
To prove it, all we do is write
_
v(t)
u(t)
f (x) dx = F (v (t)) −F (u(t))
and invoke the chain rule.
EXERCISES 111
Theorem 4.1 is one of the basic tools in the study of differential equa
tions, where the basic problemis to ﬁnd a function whose derivative satisﬁes
a given equation. While we can often recognize antiderivatives by revers
ing the formulas for derivatives, Theorem 4.1 provides the simplest general
formula for deﬁning a function in terms of its derivative.
Some problems in modern applied mathematics appear to involve
derivatives of functions that aren’t necessarily differentiable. In many
of these cases, it turns out that what we really need is the possibility of
recovering the function by integrating a nominal derivative, not the pos
sibility of producing a derivative by differentiating the original function.
We’ll see an example of such a situation when we investigate arc length. A
generalized notion of derivative, recently developed, is based on integration
instead of differentiation and is often used in such problems. Theorem 4.1
provides the ﬁrst hints toward such a theory.
EXERCISES
14. For x ∈ R, deﬁne f (x) to be the largest n ∈ Z with n ≤ x, and
deﬁne F (t) =
_
t
0
f (x) dx. Where is F continuous? Where is it
differentiable?
15. What is the largest interval on which F (t) =
_
t
0
dx
x
2
−a
2
deﬁnes F
as a continuous function?
16. Suppose that f is Riemann integrable over every closed interval [a, b],
and F is deﬁned by
F (t) =
_
t
0
__
s
0
f (x) dx
_
ds.
Explain why this formula deﬁnes F (t) for all t ∈ R, and why F is
differentiable. Where is F
differentiable?
17. For f a numerical function that is continuous on the open interval I
and a ∈ I, show that
F (t) =
_
t
a
(t −x) f (x) dx
deﬁnes a function F ∈ C
2
(I) with
F (a) = F
(a) = 0, and F
(t) = f (t) for all t ∈ I.
(Note that Theorem 4.1 doesn’t apply directly to F (t) because the
function to be integrated depends on t as well as x. Rewrite the
formula for F (t) before trying to ﬁnd F
(t).)
112 CHAPTER V THE RIEMANN INTEGRAL
18. For f a numerical function that is continuous on the open interval I
and a ∈ I, show that
G(t) =
1
2
_
t
a
(t −x)
2
f (x) dx
deﬁnes a function G ∈ C
3
(I) with
G(a) = G
(a) = G
(a) = 0, and G
(3)
(t) = f (t).
19. The two preceding exercises are special cases of a more general
formula. Try to discover it.
5 THE FUNDAMENTAL THEOREM OF CALCULUS
In the previous section we showed how we can use integration to deﬁne an
tiderivatives. Now we’ll turn the process around and show how we can use
antiderivatives to evaluate integrals. This is considered so important that
the theorembelowis usually called the fundamental theoremof calculus;
some authors also consider Theorem 4.1 to be part of the fundamental the
orem. Our versions of these theorems are slightly more complicated than
those found in basic calculus texts because we want to include the case of
discontinuous integrable functions.
Theorem 5.1: Let f be a numerical function that is Riemann inte
grable over the closed interval [a, b], and let F be a continuous numerical
function on [a, b]. Suppose that with the possible exception of ﬁnitely many
points in (a, b), F is differentiable with F
(x) = f (x). Then
_
b
a
f (x) dx = F (b) −F (a).
If we assume f and F to be continuous on [a, b] with F
(x) = f (x)
at all points of (a, b), it’s easy to prove Theorem 5.1 by using Theorem 4.1.
In that case we can deﬁne
G(t) = F (t) −
_
t
a
f (x) dx for t ∈ [a, b],
making G continuous on [a, b] with G
(x) = 0 on (a, b). Then the mean
value theorem shows that G(b) −G(a) = 0, and therefore
F (b) −
_
b
a
f (x) dx −F (a) = 0.
5. THE FUNDAMENTAL THEOREM OF CALCULUS 113
Solving for the integral gives the stated formula. However, we’ve made no
assumption of continuity for f, so a different argument is needed to prove
the fundamental theorem in the form we stated.
Proof : We’ll ﬁrst consider a special case, and then show how to
reduce the general case to the one we’ve considered. When F is differen
tiable with F
(x) = f (x) at all points in (a, b), we prove that F (b)−F (a)
is the value of the integral by showing it is in R(P) for every partition P,
making it the one number in [L(P), U (P)] for every choice of P. Once
P ={x
k
}
n
k=1
has been speciﬁed, using the mean value theorem on each
subinterval [x
k−1
, x
k
] shows there are sampling points {x
∗
k
}
n
k=1
with
f (x
∗
k
) ∆x
k
= F (x
k
) −F (x
k−1
) for k = 1, . . . , n.
Then
n
k=1
f (x
∗
k
) ∆x
k
=
n
k=1
[F (x
k
) −F (x
k−1
)]
= F (x
n
) −F (x
0
) = F (b) −F (a),
showing that F (b) −F (a) is in R(P).
For the general case of the fundamental theorem, we list the exceptional
points in (a, b) in increasing order; then we can assume
a = x
0
< x
1
< x
2
< · · · < x
n
= b
with F continuous on each subinterval [x
k−1
, x
k
] and F
(x) = f (x) for
all x ∈ (x
k−1
, x
k
). Then for each k we have
_
x
k
x
k−1
f (x) dx = F (x
k
) −F (x
k−1
),
and so
_
b
a
f (x) dx =
n
k=1
_
x
k
x
k−1
f (x) dx
=
n
k=1
[F (x
k
) −F (x
k−1
)] = F (b) −F (a).
That takes care of the general case.
In most calculus classes we’re taught to compute
_
b
a
f (x) dxby simply
ﬁnding an antiderivative F (x) and then evaluating F (b)−F (a). But there
114 CHAPTER V THE RIEMANN INTEGRAL
really are hypotheses to be checked, and they’re not always satisﬁed. For
example,
_
b
a
dx
x
2
= −
1
b
+
1
a
is only true when 0 / ∈ [a, b] because the function 1/x
2
isn’t integrable over
any interval that includes the origin. Besides checking f for integrability
over [a, b], at each point where F is not differentiable we need to check F
for continuity. Problems can arise in surprising ways. For example,
F (x) =
_ √
x
2
+ 1, x ≥ 0
−
√
x
2
+ 1, x < 0
satisﬁes
F
(x) = f (x) =
x
√
x
2
+ 1
for all x = 0,
so F
has a removable discontinuity at x = 0 even though the discontinuity
of F at 0 is not removable. While
_
b
a
f (x) dx exists for all [a, b], it’s only
given by F (b) − F (a) for 0 / ∈ (a, b]. Something deﬁnitely goes wrong
when a < 0 ≤ b, because in that case F (b) − F (a) is larger than b − a
yet all the values of f are between 0 and 1.
Although it’s unusual to use Theorem 5.1 to evaluate
_
b
a
f (x) dx over
an interval where f has a discontinuity, there are simple cases where it can
be done. For example,
f (x) =
_
1, x ≥ 0
−1, x < 0
deﬁnes a function with a discontinuity at 0, and F (x) = x deﬁnes a
continuous function with F
(x) = f (x), x = 0. In this case Theorem 5.1
says that
_
b
a
f (x) dx = b −a for all a, b ∈ R.
EXERCISES
20. Both arctanx and arctan(−1/x) are antiderivatives for
_
1 +x
2
_
−1
.
Over which intervals [a, b] can each be used to evaluate
_
b
a
dx
1 +x
2
?
6. TOPICS FOR FURTHER STUDY 115
21. Let F (x) = x −c −x −d, where c and d are real constants with
c < d. Find a function f (x) such that
_
b
a
f (x) dx = F (b) −F (a) for every closed interval [a, b].
22. The integrationbyparts formula is a consequence of the product rule
for derivatives and Theorem 5.1. State appropriate hypotheses for it.
6 TOPICS FOR FURTHER STUDY
There’s an “obvious” inequality for integrals that’s missing from this chap
ter: if b−a > 0 and f is Riemann integrable on [a, b] with m < f (x) < M
for all x ∈ [a, b], then
m(b −a) <
_
b
a
f (x) dx < M (b −a).
Since we haven’t assumed that f is continuous, that inequality is remarkably
difﬁcult to prove. Note that it implies there are subintervals where the
inﬁmum of the values of f is strictly larger than m and others where the
supremum of the values of f is strictly smaller than M. An equivalent
formulation is a little easier to deal with: if f is nonnegative and Riemann
integrable over [a, b] with
_
b
a
f (x) dx = 0, then there is a point in [a, b]
where f vanishes. One way to prove it is to showthere is a nested sequence
of closed intervals {[a
n
, b
n
]}
∞
n=1
with
sup{f (x) : a
n
≤ f (x) ≤ b
n
} < 1/n;
then f must vanish on their intersection.
With a careful application of the tools we’ve developed, it’s possible to
prove that if f is Riemann integrable on [a, b] and Φis uniformly continuous
on {f (x) : a ≤ x ≤ b}, then F (x) = Φ(f (x)) also deﬁnes a Riemann
integrable function on [a, b]. Proofs appear in many books on real analysis,
such as the one by Rudin [5].
Sometimes we integrate functions containing an extra variable that re
mains ﬁxed while x varies over the interval of integration; this happened
in some of the exercises. Leibniz observed that under fairly general hy
potheses, we might consider the integral to be a differentiable function of
the extra variable, and in those cases Leibniz’s formula
d
dy
_
b
a
F (x, y) dx =
_
b
a
F
2
(x, y) dx
is often useful.
VI
EXPONENTIAL AND
LOGARITHMIC FUNCTIONS
A
ny serious study of calculus leads to the discovery that many
problems can’t be solved by working only with functions deﬁned
by algebraic formulas. We often need functions that transcend
this limitation. Fortunately, calculus provides us with the means to produce
transcendental functions; in this chapter we’ll study the simplest of these.
1 EXPONENTS AND LOGARITHMS
Since the laws for working with exponents and logarithms are consequences
of the meanings we assign to these terms, we’ll begin by recalling howthey
are developed in algebra. The fundamental purpose of an exponent is to
indicate the number of times a factor appears in a product, which accounts
for the fundamental law of exponents:
(a
m
) (a
n
) = a
m+n
for all m, n ∈ N.
116
1. EXPONENTS AND LOGARITHMS 117
The fact that mn is the result of adding mto itself n times accounts for the
secondary rule:
(a
m
)
n
= a
mn
for all m, n ∈ N.
This exponential notation is used for all sorts of products of mathematical
objects, not just multiplication of numbers, so the two laws above appear
in a variety of contexts.
In algebra, exponential notation is extended to include exponents that
are not natural numbers. When a is a positive real number, there’s an
algebraic meaning assigned to a
r
for every rational number r. (There are
also times when fractional powers of negative numbers are used, but we
won’t consider that case here.) For a > 0, we ﬁrst extend exponential
notation to all of Z by deﬁning
a
0
= 1 and a
−n
=
1
a
n
.
After checking that the two laws of exponents remain valid for all m, n ∈ Z,
we deﬁne rational exponents by calling a
m/n
the positive solution of the
equation x
n
= a
m
when m ∈ Z and n ∈ N.
However, there’s a problem with that process because the equation
we’re solving to ﬁnd x depends on m and n separately. To deﬁne a
r
for
r ∈ Q, the value we assign to it must depend only on the value of r as a
rational number, yet there are many different ways to write a single rational
number as a fraction. To prove we have a valid deﬁnition, it’s necessary to
show that whenever m/n = m
/n
, the equations x
n
= a
m
and x
n
= a
m
have the same positive solution. That is, we need to assume that a, x, and
y are positive real numbers, that m, m
, n, and n
are natural numbers with
n and n
positive, and ﬁnally that
x
n
= a
m
and y
n
= a
m
with
m
n
=
m
n
.
We prove x = y as a consequence of those assumptions. We ﬁrst note that
y
n
n
=
_
y
n
_
n
=
_
a
m
_
n
= a
m
n
as well as
x
nn
= (x
n
)
n
= (a
m
)
n
= a
mn
.
Since the assumption m/n = m
/n
tells us that m
n = mn
, we see that
x and y are positive solutions of the same equation—an equation that has
118 CHAPTER VI EXPONENTIAL AND LOGARITHMIC FUNCTIONS
just one positive solution. Hence x = y, and our deﬁnition of rational
exponents is a valid one.
The next step is to show that the two basic laws of exponents remain
valid for arbitrary rational exponents:
(a
r
) (a
s
) = a
r+s
for all r, s ∈ Q,
(a
r
)
s
= a
rs
for all r, s ∈ Q.
Finally, showing that
(ab)
r
= a
r
b
r
and
_
a
b
_
r
=
a
r
b
r
.
whenever a and b are positive real numbers and r ∈ Q completes the
development of the algebraic theory of exponents.
In algebra, logarithms are introduced in connection with exponents.
For a and x positive numbers, it’s customary to deﬁne
log
a
(x) = r if x = a
r
.
This deﬁnition and the laws of exponents lead to the familiar laws of
logarithms. For example, if we assume that
log
a
(x) = r and log
a
(y) = s,
then since
xy = (a
r
) (a
s
) = a
r+s
we have
log
a
(xy) = r +s = log
a
(x) + log
a
(y).
However, if we regard a
r
as deﬁned only when r ∈ Q, then logarithms
become quite problematic because we lack a criterion for recognizing pos
itive real numbers as expressible in the form a
r
with r rational. That
problem did not block the development of logarithms as a computational
tool, and for several hundred years scientists found logarithms by referring
to tables and interpolating between the tabulated values. There was an
underlying assumption that log
a
(x) is a continuous, monotonic function
deﬁned for all x > 0. Of course, if log
a
(x) really is a continuous function
deﬁned on (0, ∞) then its values can’t be limited to the rational num
bers; the intermediate value theorem doesn’t allow such things. Irrational
exponents are needed to complete the theory of logarithms.
EXERCISES 119
EXERCISES
1. When a < 0 but a
m
has an nth root, this root is sometimes indicated
by a
m/n
. Show that with this interpretation, the value of (−1)
m/n
depends on mand n separately instead of depending only on the value
of m/n.
2. Prove that when a, b > 0 and r ∈ Qwe have
(ab)
r
= a
r
b
r
.
2 ALGEBRAIC LAWS AS DEFINITIONS
As we expanded our notion of exponents from Nto Q, we made sure that
the basic laws of exponents and logarithms remained valid. That remains
our goal as we expand our notion of exponents to all of R. Instead of
worrying immediately about how to deﬁne a
x
or log
a
(x) for arbitrary x,
we’ll focus our attention on understanding the properties we expect them
to have. Our ﬁrst step is to deﬁne general exponential and logarithmic
functions to be otherwise unspeciﬁed functions that satisfy appropriate
conditions. Then we will develop the properties that such functions must
have as a consequence. In the next two sections we’ll use the methods
of calculus to introduce the most important examples of exponential and
logarithmic functions: the natural exponential function and the natural
logarithm. We’ll be able to deﬁne a
x
and log
a
(x) in terms of these two
functions.
Definition 2.1: An exponential function is a differentiable func
tion E : R →(0, ∞) such that
E (x +y) = E (x) E (y) for every x, y ∈ R.
A logarithmic function is a differentiable function L : (0, ∞) →Rsuch
that
L(xy) = L(x) +L(y) for every x, y > 0.
Since 0+0 = 0 and 1· 1 = 1, it’s clear that every exponential function
satisﬁes
E (0) = 1
and every logarithmic function satisﬁes
L(1) = 0.
120 CHAPTER VI EXPONENTIAL AND LOGARITHMIC FUNCTIONS
The constant functions E
0
and L
0
deﬁned by E
0
(x) = 1 and L
0
(x) = 0
satisfy our deﬁnitions of exponential and logarithmic functions. But since
they’re not of much interest, we’ll call these the trivial exponential and
logarithmic functions. We’ll need to establish the existence of nontrivial
exponential and logarithmic functions for our deﬁnitions have any real use.
We begin by exploring the consequences of our deﬁnitions. First we
derive some additional equations that must be satisﬁed by any exponential
function E. For any n ∈ N, the fact that adding any x to itself n times
produces nx leads easily to the identity
E (nx) = [E (x)]
n
. (6.1)
A formal proof can be worked out using mathematical induction. Next,
since
1 = E (0) = E (−nx +nx) = E (−nx) E (nx),
we note that equation (6.1) is valid for all n ∈ Z as well as n ∈ N. Thus
for all m/n with m ∈ Z and n ∈ N,
[E (m/n)]
n
= E (m) = [E (1)]
m
,
so our deﬁnition of rational exponents shows that
E (r) = [E (1)]
r
for every r ∈ Q.
That’s why we called a function Ethat satisﬁes our deﬁnition an exponential
function. The number E (1) is called its base.
Now let L be an arbitrary logarithmic function. For any x > 0, we
know that
L(x
2
) = L(x · x) = L(x) +L(x) = 2L(x),
and, more generally, that
L(x
n
) = nL(x) for every n ∈ N.
Replacing x by x
1/n
shows that
L(x) = nL
_
x
1/n
_
,
and then replacing x by x
m
with m ∈ Nshows that
L
_
x
m/n
_
=
1
n
L(x
m
) =
m
n
L(x).
2. ALGEBRAIC LAWS AS DEFINITIONS 121
Since
0 = L(1) = L
_
x
0
_
= L
_
x
−m/n
x
m/n
_
= L
_
x
−m/n
_
+L
_
x
m/n
_
,
we see that
L(x
r
) = rL(x) for every x > 0 and every r ∈ Q.
When L is a nontrivial logarithmic function, its set of values must be
all of R. Given any y ∈ R, once we pick a > 0 with L(a) = 0 we can
ﬁnd a positive integer n such that nL(a) > y. Then y must be between
L(a
−n
) and L(a
n
). Since L is differentiable, it is continuous. So the
intermediate value theorem guarantees that there is an x between a
−n
and
a
n
with L(x) = y.
In particular, whenever L is a nontrivial logarithmic function there
must be a positive number b with L(b) = 1, and we can then write
L(b
r
) = r for every r ∈ Q.
That justiﬁes our calling L a logarithmic function; the number b is called
its base.
Using Deﬁnition 2.1, it’s easy to derive a formula for derivatives of
arbitrary exponential functions. First, note that if E is any exponential
function and a is any constant, then for all x ∈ Rwe have
E (x) = E (a) E (x −a).
Consequently, with the help of the chain rule we ﬁnd that
E
(x) = E (a) E
(x −a)
d
dx
(x −a) = E (a) E
(x −a).
Then setting x = a shows that
E
(a) = E (a) E
(0).
Since a was an arbitrary real number, we must have
E
(x) = E
(0) E (x) for all x ∈ R. (6.2)
Nowlet’s ﬁnd an equation satisﬁed by the derivative of any logarithmic
function L. According to Deﬁnition 2.1, for any constant a > 0 and for all
x > 0 we can write
L(x) = L
_
x
a
_
+L(a).
122 CHAPTER VI EXPONENTIAL AND LOGARITHMIC FUNCTIONS
This time we ﬁnd that
L
(x) = L
_
x
a
_
d
dx
_
x
a
_
=
1
a
L
_
x
a
_
.
Again setting x = a, we obtain
L
(a) =
1
a
L
(1).
Since a can be any positive real number, we conclude that
L
(x) =
L
(1)
x
for all x > 0. (6.3)
Next we show that exponential and logarithmic functions are charac
terized by equations (6.2) and (6.3) and their values at a single point. That’s
the content of the theorem below.
Theorem 2.1: Adifferentiable function E : R →(0, ∞) with E (0)
= 1 is an exponential function if and only if E
(x) /E (x) is constant.
A differentiable function L : (0, ∞) →Rwith L(1) = 0 is a logarith
mic function if and only if xL
(x) is constant.
Proof : We’ve already seen that E
(x) /E (x) is the constant
E
(0) whenever E is an exponential function. Now let’s assume that E
is a differentiable function with E (x) > 0 for all x ∈ R and that a is an
arbitrary constant. Then we deﬁne f : R →(0, ∞) by the formula
f (x) =
E (x +a)
E (x)
.
According to the quotient rule and the chain rule,
f
(x) =
E (x) E
(x +a) −E (x +a) E
(x)
E (x)
2
=
E (x +a)
E (x)
_
E
(x +a)
E (x +a)
−
E
(x)
E (x)
_
.
Thus if E
(x) /E (x) is constant, then f
(x) = 0 for all x. The mean
value theorem then guarantees that f is constant, and so for all x ∈ R we
have
E (x +a)
E (x)
= f (x) = f (0) =
E (a)
E (0)
.
EXERCISES 123
If we further assume that E (0) = 1, we have
E (x +a) = E (x) E (a) for all x, a ∈ R,
proving that E is an exponential function.
Before we stated the theorem, we sawthat xL
(x) is constant for every
logarithmic function L. On the other hand, if we simply assume that L is a
differentiable function on (0, ∞), then for any a > 0 we can deﬁne another
differentiable function g : (0, ∞) →Rby the formula
g (x) = L(ax) −L(x).
By the chain rule,
g
(x) = aL
(ax) −L
(x),
and so multiplication by x yields
xg
(x) = axL
(ax) −xL
(x).
Consequently, if xL
(x) is constant, then g
(x) = 0 for all x > 0. So
g (x) must be constant as well. In that case, for all x > 0 we have
L(ax) −L(x) = g (x) = g (1) = L(a) −L(1).
With the assumption that L(1) = 0, we then ﬁnd that
L(ax) = L(a) +L(x) for all a, x > 0,
proving that L is a logarithmic function.
EXERCISES
3. Prove that the product of two exponential functions is also an expo
nential function.
4. Suppose that E is a nontrivial exponential function. Find all the
differentiable functions f : R →Rsuch that the composite function
E ◦ f is also a nontrivial exponential function.
5. Suppose that L is a nontrivial logarithmic function. Find all the
differentiable functions f : R →Rsuch that the composite function
f ◦ L is also a nontrivial logarithmic function.
6. Prove that if E is a nontrivial exponential function, then it has an
inverse and its inverse is a logarithmic function.
124 CHAPTER VI EXPONENTIAL AND LOGARITHMIC FUNCTIONS
3 THE NATURAL LOGARITHM
In the last section we talked about general exponential and logarithmic
functions, but the only ones we’ve seen so far are the trivial ones. Now
we begin to remedy that situation. There’s a remarkably simple way to
produce a logarithmic function, a way that seems to have little to do with
either exponents or logarithms.
Definition 3.1: The natural logarithm is the function log :
(0, ∞) →Rdeﬁned by the formula
log (x) =
_
x
1
t
−1
dt.
Since f (t) = t
−1
deﬁnes a continuous function on (0, ∞), Theorem
4.1 in Chapter 5 shows that the natural logarithmis a differentiable function,
with its derivative given by
log
(x) = x
−1
for all x > 0.
It also satisﬁes
log (1) =
_
1
1
t
−1
dt = 0,
so Theorem 2.1 shows that the natural logarithm really is a logarithmic
function. It’s certainly not the trivial function, since its derivative is never
zero.
The base of the natural logarithm is ordinarily called e and is deﬁned
by the equation
log (e) = 1.
To estimate e, we examine values of the natural logarithm. The integral
deﬁning log (2) shows us that
1
2
≤ log (2) ≤ 1, so therefore 1 ≤ log (4) ≤
2 and it follows that
2 ≤ e ≤ 4.
Other characterizations of e have been used to prove that it is an irrational
number, and its decimal representation has been calculated to an enormous
number of places.
It isn’t terribly difﬁcult to compute the value of log x quite accurately
when x is a given number. Riemann sums provide a natural way to do this,
3. THE NATURAL LOGARITHM 125
and with the right partition and choice of sampling points they’re easy to
calculate. First let’s suppose that x > 1. Instead of using subintervals of
equal length, we partition [1, x] into n subintervals by deﬁning
t
k
= x
k/n
for k = 0, . . . , n.
For this partition, the length of the kth subinterval is
∆t
k
= x
k/n
−x
(k−1)/n
= x
(k−1)/n
_
x
1/n
−1
_
,
and so the mesh of this partition is then
x
1−1/n
_
x
1/n
−1
_
.
Since x
1/n
→ 1 as n → ∞, we see that we can make the mesh of this
partition arbitrarily small by choosing a large enough value for n. We
choose our sampling points as
t
∗
k
= t
k−1
= x
(k−1)/n
,
and then the corresponding Riemann sum is
n
k=1
1
t
∗
k
∆t
k
=
n
k=1
x
(k−1)/n
_
x
1/n
−1
¸
x
(k−1)/n
= n
_
x
1/n
−1
_
.
If 0 < x < 1 we modify the process slightly, but it isn’t really very
different. In this case
log (x) =
_
1
x
−t
−1
dt,
so our interval of integration is [x, 1] instead. We deﬁne
t
k
= x
1−k/n
for k = 0, . . . , n
so that
∆t
k
= x
1−k/n
−x
1−(k−1)/n
= x
1−k/n
_
1 −x
1/n
_
.
In this case we choose
t
∗
k
= t
k
= x
1−k/n
to obtain the same approximation that we found before.
126 CHAPTER VI EXPONENTIAL AND LOGARITHMIC FUNCTIONS
Thus for all x > 0, we obtain the formula
log (x) = lim
n→∞
n
_
x
1/n
−1
_
.
There’s even an efﬁcient way to calculate this limit by using repeated square
roots; it amounts to considering only powers of 2 for n. If we deﬁne
L
k
(x) = 2
k
_
x
1/2
k
−1
_
,
then x
1/2
k
= 1 + 2
−k
L
k
(x) and x
1/2
k+1
is its square root. Consequently,
L
k+1
(x) = 2
k+1
_
_
1 + 2
−k
L
k
(x) −1
_
,
so successive approximations can be calculated quite easily. The calcula
tions begin with
L
1
(x) = 2
_√
x −1
¸
.
A better way to do the calculations is to rewrite
L
k+1
(x) = 2
k+1
_
1 + 2
−k
L
k
(x) −1
_
1 + 2
−k
L
k
(x) + 1
_
=
2L
k
(x)
1 +
_
1 + 2
−k
L
k
(x)
.
As k → ∞,
_
1 + 2
−k
L
k
(x) becomes increasingly difﬁcult to distinguish
from1, and eventually both digital computers and calculators will treat them
as exactly equal. At that stage, the ﬁrst formula returns L
k+1
(x) = 0, but
the second returns L
k+1
(x) = L
k
(x). Unless x is either extremely large
or extremely close to 0, this stage can be reached in a reasonable number
of steps.
EXERCISES
7. Prove that if 1 ≤ x < y, then log y − log x < y − x. Does the last
inequality remain true for 0 < x < y ≤ 1 or for 0 < x < 1 < y?
8. By writing log (f (x)) as
1
2
log
_
[f (x)]
2
_
, show that when f (x) is
a nonzero differentiable function we always have
d
dx
log (f (x)) =
f
(x)
f (x)
.
9. Show that for x > 1 the sequence {L
k
(x)}
∞
k=1
is a decreasing se
quence with positive terms and that for 0 < x < 1 it is a decreasing
sequence with negative terms.
10. Show that e > 2.5 by ﬁnding an integer k such that L
k
(2.5) < 1.
4. THE NATURAL EXPONENTIAL FUNCTION 127
4 THE NATURAL EXPONENTIAL FUNCTION
While we’ve characterized an exponential function as a differentiable func
tion E : R →(0, ∞) such that E (0) = 1 and E
(x) /E (x) is constant, we
haven’t yet produced a nontrivial one. There are quite a few processes that
produce them, but with most of these methods it’s very difﬁcult to prove
that the function they produce really is an exponential function. There is
one notable exception, and that’s the route we’ll follow. We’ll show that
the inverse of the natural logarithm is an exponential function, commonly
known as the natural exponential function. But deﬁning a function as
the inverse of a function whose values we can only approximate isn’t very
satisfying. So in the next section we’ll also develop a method for express
ing the value of the natural exponential function directly as the limit of
a sequence of approximations, with each approximation involving only a
ﬁnite number of arithmetic operations.
Definition 4.1: The inverse of the natural logarithm is called the
natural exponential function. We denote the natural exponential function
by exp, so that
y = log (x) if and only if x = exp(y).
Since the natural logarithm has domain (0, ∞) with no critical points
and its set of values is all of R, the inverse function theorem tells us that
exp : R →(0, ∞) with
exp
(y) =
1
log
(x)
= x = exp(y)
when y = log (x). Hence exp
(y) / exp(y) is indeed constant. We also
know exp(0) = 1 because log (1) = 0, so the inverse of the natural
logarithm satisﬁes our deﬁnition of exponential function.
The natural logarithm and the natural exponential function give us an
easy way to interpret irrational exponents. Since
log (a
r
) = r log (a) when a > 0 and r ∈ Q,
for all such a and r we have
a
r
= exp[r log (a)].
We simply agree that
a
x
= exp[xlog (a)] for a > 0 and x ∈ R.
128 CHAPTER VI EXPONENTIAL AND LOGARITHMIC FUNCTIONS
This agrees with the algebraic deﬁnition of exponents when x is rational,
and since
d
dx
(a
x
) =
d
dx
{exp[xlog (a)]}
= exp
[xlog (a)]
d
dx
[xlog (a)]
= exp[xlog (a)] log (a)
= a
x
log (a),
we see that it is indeed an exponential function. The usual laws of exponents
are then valid with arbitrary real exponents:
(a
x
) (a
y
) = exp[xlog (a)] exp[y log (a)]
= exp[xlog (a) +y log (a)]
= exp[(x +y) log (a)] = a
x+y
as well as
(a
x
)
y
= exp[y log (a
x
)] = exp[xy log (a)] = a
xy
.
This extension of the deﬁnition of a
x
to the case of arbitrary exponents
resolves the difﬁculty with the deﬁnition of log
a
(x). We can again deﬁne
log
a
(x) = y if and only if x = a
y
,
and since
log (a
y
) = y log (a)
we see that
log
a
(x) =
log (x)
log (a)
.
That makes
d
dx
[log
a
(x)] =
1
log (a)
d
dx
[log (x)] =
1
xlog (a)
.
Consequently, we see that this version of log
a
satisﬁes our deﬁnition of
logarithmic function.
One important application of our deﬁnition of a
x
is the case a = e, the
base of the natural logarithm. Since e is deﬁned by the condition
log (e) = 1,
EXERCISES 129
we see that
e
x
= exp[xlog (e)] = exp(x).
Consequently, the natural exponential function is often expressed in expo
nential notation rather than functional notation.
EXERCISES
11. Why didn’t we use the equation E
(x) = E (x) E
(0) to deﬁne
exp(x) directly, in much the same way we deﬁned log (x)?
12. Prove that (ab)
x
= (a
x
) (b
x
) for all a, b > 0 and all x ∈ R.
13. Show that if E is any exponential function, then log [E (x)] has a
constant derivative. Then explain why every exponential function
can be expressed in the formE (x) = exp(ax) for some constant a.
14. Suppose that f and g are differentiable functions, with f positive.
Show that
d
dx
_
[f (x)]
g(x)
_
= [f (x)]
g(x)
_
f
(x)
g (x)
f (x)
+g
(x) log [f (x)]
_
.
15. The hyperbolic sine and hyperbolic cosine are deﬁned by the formulas
sinht =
1
2
_
e
t
−e
−t
_
and cosh t =
1
2
_
e
t
+e
−t
_
.
Show that each point on the right half of the hyperbola x
2
−y
2
= 1
has the form (cosh t, sinht) for exactly one t ∈ R. Then show that
d
dt
(sinht) = cosht and
d
dt
(cosh t) = sinht.
5 AN IMPORTANT LIMIT
We saw that the natural logarithm can be computed from the formula
log (x) = lim
n→∞
n
_
x
1/n
−1
_
for any x > 0. Consequently, for any x ∈ Rwe have
lim
n→∞
n
_
e
x/n
−1
_
= log (e
x
) = x.
This says that as n increases, the difference between e
x/n
and 1 +
1
n
x
becomes negligible in comparison to 1/n. Since
e
x
=
_
e
x/n
_
n
,
130 CHAPTER VI EXPONENTIAL AND LOGARITHMIC FUNCTIONS
can we say that
e
x
= lim
n→∞
_
1 +
1
n
x
_
n
?
The answer is an emphatic yes, and there are many ways to prove it.
We’ll prove a more general result:
lim
t→0
(1 +at)
1/t
= e
a
for all a ∈ R.
We begin by noting that
f (t) = log (1 +at)
deﬁnes a function that is differentiable at t = 0, and
f
(t) =
1
1 +at
·
d
dt
(at) =
a
1 +at
for 1 +at > 0.
Since f
(0) = a, the deﬁnition of the derivative tells us that
a = lim
t→0
f (t) −f (0)
t −0
= lim
t→0
log (1 +at)
t
.
We know that the natural exponential function is continuous, so
exp(a) = lim
t→0
exp
_
log (1 +at)
t
_
= lim
t→0
(1 +at)
1/t
.
The limit of the sequence
__
1 +
1
n
x
_
n
_
∞
n=1
is sometimes used to deﬁne
e
x
. It really is possible to prove that this sequence converges for each x and
that its limit deﬁnes an exponential function. But it’s a lengthy process,
and we won’t begin it; one deﬁnition of the natural exponential function
is enough. However, we will examine the sequence further and derive an
estimate for the error in approximating e
x
by the polynomial
_
1 +
1
n
x
_
n
.
Since e
x
=
_
e
x/n
_
n
, the mean value theorem tells us that
e
x
−
_
1 +
1
n
x
_
n
= nξ
n−1
_
e
x/n
−
_
1 +
1
n
x
__
(6.4)
for some ξ between
_
1 +
1
n
x
_
and e
x/n
. Since
exp
(0) = exp(0) = 1,
EXERCISES 131
we recognize that 1 +
1
n
x is just the linearization of exp near 0, evaluated
at x/n instead of x. That suggests using Theorem 4.1 in Chapter 4 to
estimate the last factor in equation (6.4). We can write
e
x/n
= 1 +
1
n
x +
1
2
_
x
n
_
2
exp(η)
for some real number η between 0 and x/n. Substituting this into (6.4)
gives us the representation
e
x
−
_
1 +
1
n
x
_
n
=
1
2n
x
2
ξ
n−1
exp(η).
Since exp(η) is between 1 and e
x/n
, it’s reasonable to estimate both ξ and
exp(η) by e
x/n
as n → ∞. We then obtain
e
x
−
_
1 +
1
n
x
_
n
≈
1
2n
x
2
e
x
as n → ∞. (6.5)
In other words, the relative error in this approximation to e
x
is about x
2
/2n
for large values of n. Unless x is quite close to 0, extremely large values
of n will be needed to produce high levels of accuracy.
EXERCISES
16. In calculating compound interest, interest rates are stated in terms of
a nominal annual interest rate r. If the year is divided into n equal
interest periods, the interest rate per interest period is r/n. Calculate
the total interest paid over a year on a unit investment, and showthat it
approaches e
r
−1 as the number of interest periods increases without
bound. This limiting case is usually called continuous compound
ing.
17. Prove that lim
x→∞
x
n
/e
x
= 0 for each n ∈ N.
18. To ten signiﬁcant ﬁgures, e = 2.718281828. Find
_
1 +
1
n
_
n
for
n = 10 and n = 100, and compare the error to the error estimate for
e
1
given by (6.5).
19. Use a quadratic approximation to the natural exponential function to
show that for all x > 0 and each positive integer n,
n
_
x
1/n
−1
_
= log (x) +
1
2n
ξ [log (x)]
2
for some ξ between 1 and x
1/n
. Note x
1/n
= exp
_
1
n
log (x)
¸
.
VII
CURVES AND ARC LENGTH
I
n the previous chapter we used calculus methods to produce functions
of a decidedly algebraic nature. Here we’ll see that the methods of
calculus can also be used to develop some of the basic functions that
we use in geometry and trigonometry to ﬁnd distances.
1 THE CONCEPT OF ARC LENGTH
Clearly the phrase arc length of a curve refers to some sort of one
dimensional measurement. But just what is a curve, and what do we
mean by its arc length? We’ll try to answer these questions by looking at
some simple geometric ideas.
If our primary interest in curves is in determining the length of a path
that follows one, then we see that it’s entirely reasonable for curves to be
straight or to have corners, and the mathematical deﬁnition of a curve does
indeed allowfor those possibilities. We normally think of a curve as linking
its endpoints in a continuous path, with the points on the curve arranged
132
1. THE CONCEPT OF ARC LENGTH 133
in a deﬁnite order. A natural way to specify this order is to represent the
curve in the form
x = f (t), y = g (t) for a ≤ t ≤ b,
with f and g continuous numerical functions on [a, b]. This is called a
parametric representation of the curve; the auxiliary variable t is called
a parameter. Given such a representation, we can write
P (t) = (f (t), g (t))
for a generic point on the curve. That orders the points: P (t
1
) precedes
P (t
2
) if t
1
and t
2
are in [a, b] with t
1
< t
2
. The ﬁrst point in the curve is
P (a), and the last point is P (b).
Specifying a curve in this manner gives us a standard way to construct
simple approximate paths. Suppose we partition [a, b] as we did in deﬁning
Riemann sums:
a = t
0
< t
1
< t
2
< · · · < t
n
= b.
Calling P
k
the point P (t
k
), we can approximate the curve by the broken
line path
P
0
P
1
, P
1
P
2
, . . . , P
n−1
P
n
formed by connecting the successively numbered points with line segments.
For any given any ε > 0, Theorem 5.2 in Chapter 2 shows that it’s possible
to choose the partition in such a way that neither f (t) nor g (t) will change
by more than ε over any of the subintervals. So we can choose a brokenline
path that stays arbitrarily close to the curve.
When we subdivide a curve into segments, the arc length of the curve
should be the same as the sum of the arc lengths of the segments, and
the arc length of each segment must be at least the distance between its
endpoints. So if we call PQ the distance between the points P and Q,
the arc length of the curve described above must be at least
P
0
P
1
 +P
1
P
2
 +· · · +P
n−1
P
n

for any partition of [a, b], and the sum is exactly the distance along the
brokenline path linking P
0
, P
1
, . . . , P
n
in that order. Since we can ap
proximate general curves by brokenline paths, it’s natural to use these
sums to deﬁne arc length.
It doesn’t take long to make an unpleasant discovery—we can’t always
deﬁne arc length. For example, let’s consider a simple rectangular spiral
134 CHAPTER VII CURVES AND ARC LENGTH
that alternates horizontal and vertical segments, with the length of the nth
segment given by
s
n
=
√
n −
√
n −1.
The length of the ﬁrst segment is 1, and the lengths decrease monotonically
to 0 as n increases. Drawing the whole curve is problematic, but Figure
7.1 shows a sketch of the ﬁrst ten segments. It’s easy to calculate that the
sum of the lengths of the ﬁrst n segments is exactly
√
n, so the entire path
can’t have a ﬁnite length, even though it spirals in to a single point without
ever crossing itself.
Figure 7.1 A rectangular spiral.
Now we do the sort of thing mathematicians always do in these situa
tions. Along with our formal deﬁnition of arc length, we deﬁne a restricted
class of curves excluding all those curves that can’t have an arc length.
Definition 1.1: A curve is said to be rectiﬁable if there is a
ﬁxed number that is never exceeded by the sum of the distances between
successive points on the curve, no matter how many points are considered.
The arc length of a rectiﬁable curve C is
L = sup
_
n
k=1
P
k−1
P
k
 : P
0
, P
1
, . . . , P
n
successive points on C
_
.
When we say that P
0
, P
1
, . . . , P
n
are successive points on C, that
means the points are on C and are encountered in that order; of course there
will always be additional points between them. For the purposes of our
deﬁnition, it’s vital that we keep the points in the same order as they appear
on C. It is quite possible for the same set of points to be retraced several
times by a rectiﬁable curve. With our deﬁnition, each retracing adds to the
arc length.
1. THE CONCEPT OF ARC LENGTH 135
Let’s see how the usual notion of the circumference of a circle matches
up with the deﬁnition of arc length we’ve just given. Classically, the
circumference of a circle can be approximated frombelowby the perimeter
of inscribed polygons and from above by the perimeter of circumscribed
ones. For example, Archimedes used that approach to determine π fairly
accurately. We would like to show that circles are rectiﬁable curves and
that the circumference of a circle is the same as its arc length.
We begin by examining sums of distances between successive points
on a single tracing of a circle. Since we’re interested in making such sums
as large as possible, we might as well assume that the last point selected is
the same as the ﬁrst, so that we’ve gone completely around the circle. In
that case, P
0
= P
n
, the line segments P
0
P
1
, P
1
P
2
, . . . , P
n−1
P
n
are the
sides of an inscribed polygon, and the sum of their lengths is its perimeter.
So we can prove the circle is a rectiﬁable curve if we can ﬁnd an upper
bound for the perimeters of all inscribed polygons.
In fact, the perimeter of a polygon inscribed in a circle is never more
than the perimeter of any circumscribed polygon. Although that seems
obvious, it’s actually fairly subtle, because it’s hard to recognize the reasons
it seems obvious to us. Fortunately, it’s easy to compare each side of an
inscribed polygon to part of the perimeter of an circumscribed one. Let’s
say that P
k−1
and P
k
are successive vertices on the inscribed polygon, and
Ois the center of the circle. We drawrays fromOthrough P
k−1
and P
k
and
extend them until they meet the circumscribed polygon; call those points
Q
k−1
and Q
k
. Because the angles OP
k
P
k−1
and OP
k−1
P
k
are equal and
therefore acute, it follows that
P
k−1
P
k
 < Q
k−1
Q
k
 unless Q
k−1
= P
k−1
and Q
k
= P
k
,
and Q
k−1
Q
k
 is never more than the length of any path from Q
k−1
to Q
k
.
That proves that any circle is indeed a rectiﬁable curve; its arc length
is the supremum of the set of perimeters of inscribed polygons and is
never more than the perimeter of any circumscribed polygon. What about
approximating the arc length by the perimeters of circumscribed polygons?
To see how that works, let’s suppose that our circle has radius r and
we’ve inscribed a regular polygon with n sides and perimeter p. By the
Pythagorean theorem, the distance from the center of the circle to each
side is
√
r
2
−s
2
, where s is half the length of a single side. So expanding
the inscribed polygon by a factor of r/
√
r
2
−s
2
produces a circumscribed
polygon. That expansion increases the perimeter by the same factor, and
the factor approaches 1 as the number of sides in the regular polygon
increases. Consequently, as the number of sides increases, the perimeters
136 CHAPTER VII CURVES AND ARC LENGTH
of both inscribed and circumscribed regular polygons do indeed approach
the arc length of the circle.
Now let’s return to the case of general rectiﬁable curves and make a
few observations about the sums we use to approximate their arc length.
Given a collection of successive points on a curve, we can always expand
it by ﬁtting an additional point into its proper place. If that happens to be at
either end of the list, it simply adds in another distance. Otherwise, one of
the distances gets replaced by the sum of two, and that sum is at least as big
as the distance it replaces. Consequently, expanding the list of successive
points on a curve can only increase the sum of the distances between them.
That’s a good thing to remember, as it plays a crucial role in the theorem
below. Note that the theorem requires a careful interpretation if the curve
passes through any point more than once, but the interpretation is a natural
one for curves given parametrically.
Theorem 1.1: Suppose that P is a point on a curve C, and that C
1
and C
2
are the curves formed by the points on C before and after P. Then
C is rectiﬁable if and only if both C
1
and C
2
are rectiﬁable, and in that case
the arc length of C is the sum of the arc lengths of C
1
and C
2
.
Proof : First assume that C is rectiﬁable, and call L its arc
length. Suppose we have successive points P
0
, P
1
, . . . , P
n
on C
1
and
Q
0
, Q
1
, . . . , Q
m
on C
2
; we need upper bounds for
n
k=1
P
k−1
P
k
 and
m
k=1
Q
k−1
Q
k
 to prove that C
1
and C
2
are rectiﬁable. Since adding P to
either collection of successive points can only increase the corresponding
sum, the only case we really need to consider is when P
n
= P = Q
0
. In
that case P
0
, P
1
, . . . , P
n
, Q
1
, Q
2
, . . . , Q
m
are successive points on C, and
since P
n
= Q
0
we see that
n
k=1
P
k−1
P
k
 +
m
k=1
Q
k−1
Q
k
 ≤ L.
Rewriting this inequality in the equivalent form
n
k=0
P
k−1
P
k
 ≤ L −
m
k=0
Q
k−1
Q
k

shows that C
1
is rectiﬁable, with arc length
L
1
≤ L −
m
k=0
Q
k−1
Q
k
.
1. THE CONCEPT OF ARC LENGTH 137
Rewriting this as
m
k=0
Q
k−1
Q
k
 ≤ L −L
1
shows that C
2
is also rectiﬁable, with arc length
L
2
≤ L −L
1
.
That proves half of the theorem.
For the other half, we assume that both C
1
and C
2
are rectiﬁable with
arc lengths L
1
and L
2
. Then for P
0
, P
1
, . . . , P
n
successive points on C, we
must have
n
k=0
P
k−1
P
k
 ≤ L
1
+L
2
.
When P is one of the points in P
0
, P
1
, . . . , P
n
we can split the sum into
two parts, with one bounded by L
1
and the other by L
2
. When P isn’t one
of them, we simply insert P in its proper place and the increased sum has
the desired bound. Thus C is rectiﬁable, with arc length L ≤ L
1
+ L
2
.
Combining the two inequalities we’ve proved for L, L
1
, and L
2
shows
that L = L
1
+L
2
, and that completes the proof.
When we start cutting a rectiﬁable curve into pieces, it’s convenient to
have a function that describes the result. For a rectiﬁable curve C deﬁned
by a pair of continuous numerical functions on [a, b], we can deﬁne an
arclength function λ on [a, b] by calling λ(t) the arc length of the part of
C between P (a) and P (t). Clearly λ is an increasing numerical function
on [a, b], and by the last theorem λ(s) −λ(t) is the arc length of the
part of C between P (s) and P (t). Such an arc length function is always
continuous on [a, b]; that’s important enough that we’ll state it as a formal
theorem.
Theorem 1.2: Let f and g be continuous numerical functions on
[a, b], and let C be the curve deﬁned by
P (t) = (f (t), g (t)) for a ≤ t ≤ b.
If C is rectiﬁable, then the length of the portion of C between P (a) and
P (t) deﬁnes a continuous numerical function of t on [a, b].
Proof : Calling the arclength function λ, we’ll show that for any
given ε > 0 there is a δ > 0 such that
λ(s) −λ(t) < ε for all s, t ∈ [a, b] with s −t < δ.
138 CHAPTER VII CURVES AND ARC LENGTH
Since λ is an increasing function, we can accomplish this by producing a
partition P = {t
0
, t
1
, . . . .t
n
} of [a, b] with the property that
λ(t
j
) −λ(t
j−1
) <
1
2
ε for j = 1, . . . , n.
Then we’ll know λ(s) −λ(t) < ε whenever s and t are in either the
same subinterval or adjacent subintervals deﬁned by P, and so our δ can
be the length of the shortest subinterval.
For any choice of P and any j, the error in approximating λ(t
j
) −
λ(t
j−1
) by P
j−1
P
j
 is never greater than the error in approximating the
arc length of all of C by
n
k=1
P
k−1
P
k
, and so
λ(t
j
) −λ(t
j−1
) ≤ P
j−1
P
j
 +
_
λ(b) −
n
k=1
P
k−1
P
k

_
.
Consequently, all we need to do is make sure that every segment of the
brokenline path is shorter than
1
4
ε and that the total length of the broken
line path corresponding to P is within
1
4
ε of the arc length of C.
By Theorem 5.2 in Chapter 2, there is a δ
0
> 0 such that every s and t
in [a, b] with s −t < δ
0
satisfy
f (s) −f (t) <
1
8
ε and g (s) −g (t) <
1
8
ε.
So whenever the mesh of P is less than δ
0
, every segment satisﬁes
P
k−1
P
k
 =
_
[f (t
k
) −f (t
k−1
)]
2
+ [g (t
k
) −g (t
k−1
)]
2
<
1
4
ε.
The deﬁnition of arc length shows there is a partition with
λ(b) <
1
4
ε +
n
k=1
P
k−1
P
k
,
and if necessary we can insert additional points to produce such a partition
with mesh smaller than δ
0
.
EXERCISES
1. We’ve proved earlier that an increasing function deﬁned on an interval
is continuous if its set of values is an interval. Why didn’t we use that
to prove Theorem 1.2?
2. ARC LENGTH AND INTEGRATION 139
2. Show that if f and g are continuous, monotonic numerical functions
on [a, b], then the curve deﬁned by
P (t) = (f (t), g (t)) for a ≤ t ≤ b
is rectiﬁable with arc length at most f (b) −f (a) +g (b) −g (a).
3. Explain why considering partitions {t
0
, t
1
, . . . , t
n
} of [a, b] with all
terms P
k−1
P
k
 arbitrarily small won’t necessarily produce sums that
approach the arc length of the curve deﬁned by P (t), even when the
curve is rectiﬁable.
4. Suppose that the curve given by P (t) for a ≤ t ≤ b has arc length
(1 +ε) P (a) P (b), where ε is a small positive number. For a ≤ t ≤
b, how far can the point P (t) be from the line segment P (a) P (b)?
2 ARC LENGTH AND INTEGRATION
The arclength function in the previous section has much in common with
functions deﬁned by integrals. In both cases we can approximate their
values in terms of ﬁnite sums associated with partitions, and reﬁning the
partitions improves the accuracy of the approximations. This isn’t just a
coincidence; it is often possible to use an integral to calculate arc length,
with the integrand derived from the way we represent the curve. The stan
dard formula for ﬁnding the arc length of a curve given parametrically by
is
P (t) = (f (t), g (t)) for a ≤ t ≤ b
L =
_
b
a
_
f
(t)
2
+g
(t)
2
dt.
This has an appealing physical interpretation when we think of the pa
rameter t as time. Then the derivatives f
(t) and g
(t) are the x and
ycomponents of the velocity of the moving point, and the integrand
v (t) =
_
f
(t)
2
+g
(t)
2
represents the magnitude of the velocity, better known as the speed. Thus
the integral formula for arc length says that we integrate speed with respect
to time to ﬁnd the distance traveled. But there are some very real restrictions
on the functions f and g that are not always met, even for very simple
curves. Our task in studying this formula is twofold: we need conditions
that assure the integrand is Riemann integrable, and we need to show that
the value of the integral is indeed the arc length.
140 CHAPTER VII CURVES AND ARC LENGTH
While we think of f
and g
as derivatives, what we really need from
them is that their integrals give us f and g:
f (t) = f (a) +
_
t
a
f
(τ) dτ and g (t) = g (a) +
_
t
a
g
(τ) dτ.
(7.1)
Of course, that’s the case when f and g are differentiable at all points
in [a, b] with f
and g
continuous, but (7.1) may be valid under more
general conditions. The possible existence of points at which f or g
is not differentiable is of no concern to us; we just need f
and g
to
be Riemann integrable functions that satisfy (7.1). Once again calling
P
k
= (f (t
k
), g (t
k
)), we calculate
P
k−1
P
k
 =
_
[f (t
k
) −f (t
k−1
)]
2
+ [g (t
k
) −g (t
k−1
)]
2
_
1/2
=
_
_
_
_
_
t
k
t
k−1
f
(t) dt
_
2
+
_
_
t
k
t
k−1
g
(t) dt
_
2
_
_
_
1/2
.
Thus the values of f
and g
determine how much the position of P (t)
can change over various subintervals of [a, b]. In particular, for f
and g
Riemann integrable over [a, b] they must be bounded, and that lets us write
P
k−1
P
k
 ≤
_
A
2
+B
2
(t
k
−t
k−1
)
when f
(t) ≤ A and g
(t) ≤ B for all t ∈ [a, b]. Such a bound shows
that C is rectiﬁable, with arc length at most
√
A
2
+B
2
(b −a).
Knowing that C is rectiﬁable simpliﬁes our task; when f
and g
are
Riemann integrable we can use Theorem 2.1 from Chapter 5 to show that v
is Riemann integrable over [a, b] and that the arc length L of C is the value
of
_
b
a
v (t) dt. That involves showing that for any ε > 0, there is a partition
of [a, b] such that every associated Riemann sum for v is within ε of L. As
usual, we let P ={t
k
}
n
k=0
represent an arbitrary partition of [a, b], and we
let {t
∗
k
}
n
k=1
represent an arbitrary set of sampling points for P. We then
write
n
k=1
v (t
∗
k
) ∆t
k
=
n
k=1
P
k−1
P
k
 +
n
k=1
(v (t
∗
k
) ∆t
k
−P
k−1
P
k
). (7.2)
Since C is rectiﬁable, we knowwe can make
n
k=1
P
k−1
P
k
 be within ε/2
of L by choosing P appropriately, so our job is to show that we can make
2. ARC LENGTH AND INTEGRATION 141
the last sum in equation (7.2) be smaller than ε/2 as well. That requires a
careful use of the Riemann integrability of f
and g
.
To help us make better use of these assumptions, let’s call m
k
and M
k
the inﬁmum and supremum of the values of f
over [t
k−1
, t
k
], so that
L(P) =
n
k=1
m
k
∆t
k
and U (P) =
n
k=1
M
k
∆t
k
are the lower and upper sums for f
relative to P. Then the diameter of the
set of Riemann sums for f
associated with the partition P is
U (P) −L(P) =
n
k=1
(M
k
−m
k
) ∆t
k
.
Let’s also write m
k
, M
k
, L
(P), and U
(P) for the analogous quantities
when f
is replaced by g
.
For each k we have
m
k
∆t
k
≤
_
t
k
t
k−1
f
(t) dt ≤ M
k
∆t
k
,
so that
ξ
k
=
1
∆t
k
_
t
k
t
k−1
f
(t) dt =
f (t
k
) −f (t
k−1
)
∆t
k
deﬁnes a number in [m
k
, M
k
]. Similarly,
η
k
=
g (t
k
) −g (t
k−1
)
∆t
k
is a number in [m
k
, M
k
], and we can write
P
k−1
P
k
 =
_
[f (t
k
) −f (t
k−1
)]
2
+ [g (t
k
) −g (t
k−1
)]
2
_
1/2
=
_
ξ
2
k
+η
2
k
∆t
k
.
To bound the difference between P
k−1
P
k
 and v (t
∗
k
) ∆t
k
, we use an
elementary inequality:
¸
¸
¸
_
a
2
+b
2
−
_
c
2
+d
2
¸
¸
¸ ≤ a −c +b −d. (7.3)
142 CHAPTER VII CURVES AND ARC LENGTH
To explain this inequality, call
O = (0, 0), P = (a, b), Q = (c, d), and R = (c, b).
The triangle inequality shows that
OP −OQ ≤ PQ ≤ PR +RQ,
and substituting for the lengths of these segments gives (7.3). So for each
k we have
v (t
∗
k
) ∆t
k
−P
k−1
P
k
 =
¸
¸
¸
¸
_
f
_
t
∗
k
_
2
+g
_
t
∗
k
_
2
−
_
ξ
2
k
+η
2
k
¸
¸
¸
¸
∆t
k
≤
_¸
¸
f
(t
∗
k
) −ξ
k
¸
¸
+
¸
¸
g
(t
∗
k
) −η
k
¸
¸
_
∆t
k
≤
_
(M
k
−m
k
) +
_
M
k
−m
k
_¸
∆t
k
.
Now we bound the last sum in (7.2). Since the absolute value of a sum
is never more than the sum of the absolute values, we get
¸
¸
¸
¸
¸
n
k=1
(v (t
∗
k
) ∆t
k
−P
k−1
P
k
)
¸
¸
¸
¸
¸
≤
n
k=1
_
(M
k
−m
k
) +
_
M
k
−m
k
_¸
∆t
k
= U (P) −L(P) +U
(P) −L
(P).
That too can be made smaller than ε/2 by choosing P appropriately, and
the same choice of P can also make
n
k=1
P
k−1
P
k
 be within ε/2 of L.
The theorem below summarizes what we’ve shown.
Theorem 2.1: Let C be the curve P (t) = (f (t), g (t)) for a ≤ t ≤
b, and suppose that f
and g
are Riemann integrable functions with (7.1)
satisﬁed on [a, b]. Then C is rectiﬁable, and its arc length is
L =
_
b
a
_
f
(t)
2
+g
(t)
2
dt.
In particular, these conditions are satisﬁed when the derivatives of f and g
are continuous on [a, b].
As we use this theorem, we should keep in mind that it gives a sufﬁcient
condition for a curve to be rectiﬁable, not a necessary one. Many different
pairs of functions can describe exactly the same curve, with the points
traced in the same order, and the theorem may only apply to some of these
representations. For example, the straight line segment
f (t) = g (t) = t, 0 ≤ t ≤ 1
EXERCISES 143
is also given by
f (t) = g (t) =
√
t, 0 ≤ t ≤ 1;
the theorem applies to the ﬁrst parametrization but not to the second.
However, when two different parametrizations satisfy the hypotheses of
the theorem, the resulting integrals must all have the same value, even
though the integrands can be quite different.
EXERCISES
5. Parametrize the top of the unit circle by calling (f (t), g (t)) the point
where the line segment from (0, 0) to (t, 1) crosses the unit circle.
Find the function v we integrate to ﬁnd arc length fromP (a) to P (b);
_
v (t) dt can be reduced to a standard form. Does some choice of a
and b make
_
b
a
v (t) dt correspond to the arc length of the entire top
half of the circle?
6. Parametrize the unit circle by calling (f (t), g (t)) the point above
the line y = −1 where the line through (0, −1) and (t, 0) intersects
the unit circle. Find the function v (t) we integrate to ﬁnd arc length
from P (a) to P (b) on the unit circle. (This parametrization is
especially useful because f, g, and v are simple rational functions.)
Does
_
b
a
v (t) dt ever correspond to the arc length of the top of the
circle?
3 ARC LENGTH AS A PARAMETER
Now let’s consider a slightly different class of curves. We again represent
C in the form
P (t) = (f (t), g (t)) for t ∈ I, with f, g ∈ C (I),
but this time we assume that I is an open interval, so that C has no ﬁrst or last
point. Instead of assuming that the entire curve is rectiﬁable, we suppose
that for every closed subinterval [a, b] ⊂ I, the portion of C between P (a)
and P (b) is rectiﬁable. Then there is still an increasing function λ deﬁned
on I with the property that λ(b) −λ(a) gives the arc length of the portion
of C between P (a) and P (b), even though λ(t) need not be the arc length
of the portion of C preceding P (t). We simply pick a base point t
0
∈ I
and specify λ(t
0
) = 0. Then λ(t) is positive for t > t
0
and negative for
t < t
0
; it’s common to call λ(t) the directed arc length. Instead of using
t to identify different points on C, in principle we can use λ(t) just as well.
If f and g are functions in C
1
(I), then all the assumptions we made
in the previous section about the functions deﬁning C are satisﬁed on each
144 CHAPTER VII CURVES AND ARC LENGTH
closed subinterval of I, so
λ(t) =
_
t
t
0
_
f
(τ)
2
+g
(τ)
2
dτ. (7.4)
Since λ(t) is the integral of a function that is continuous on I, Theorem
4.1 in Chapter 5 tells us that
λ
(t) =
_
f
(t)
2
+g
(t)
2
for all t ∈ I.
Thus λ is itself a function in C
1
(I). If we also assume that f
(t) and g
(t)
do not vanish simultaneously, then λ
(t) > 0 for all t ∈ I. In that case,
the inverse function theorem shows that the equation
s = λ(t)
deﬁnes t as a differentiable function of s. That is, there is a differentiable
function T on an open interval J such that
T (λ(t)) = t for all t ∈ I,
and its derivative satisﬁes
T
(s) =
1
λ
(t)
when s = λ(t).
Of all the ways to describe the same curve, mathematicians favor
expressing x and y as functions of the directed arc length. The reason is
that the parameter s has an intrinsic geometric meaning determined by the
points on C and the order in which they are encountered. Consequently, the
derivatives
dx
ds
and
dy
ds
reﬂect the geometry of the curve in ways that f
(t)
and g
(t) may not. In principle, other parametric representations can be
converted to this form by ﬁnding T (s) and substituting it for t:
x = f (T (s)), y = g (T (s)) for s ∈ J.
In practice, we may not be able to ﬁnd a usable formula for T (s), but we
can still ﬁnd the derivatives
dx
ds
and
dy
ds
in terms of t by using the chain rule
and our formula for T
(s).
For example, let’s consider a fairly arbitrary curve that follows the unit
circle; that is, a curve described by
x = f (t), y = g (t) with f (t)
2
+g (t)
2
= 1 for all t ∈ I.
3. ARC LENGTH AS A PARAMETER 145
To be able to use our formulas, we’ll assume that f and g are in C
1
(I), and
we’ll further assume that
f (t) g
(t) −f
(t) g (t) > 0 for all t ∈ I.
In addition to implying that f
and g
never vanish simultaneously, this also
implies that the motion is counterclockwise because it makes g (t) /f (t)
increasing where f (t) = 0. We’ll calculate
dx
ds
and
dy
ds
for this curve. By
the way, all our assumptions are valid for
f (t) =
1 −t
2
1 +t
2
and g (t) =
2t
1 +t
2
,
but our calculations are equally valid for many other parametrizations.
By differentiating the equation
f (t)
2
+g (t)
2
= 1
with respect to t, we quickly discover that
f (t) f
(t) = −g (t) g
(t).
Consequently, we may write
f
(t) =
_
f (t)
2
+g (t)
2
_
f
(t)
= f (t)
_
−g (t) g
(t)
¸
+g (t)
2
f
(t)
= −
_
f (t) g
(t) −f
(t) g (t)
¸
g (t)
as well as
g
(t) =
_
f (t)
2
+g (t)
2
_
g
(t)
= f (t)
2
g
(t) +g (t)
_
−f (t) f
(t)
¸
=
_
f (t) g
(t) −f
(t) g (t)
¸
f (t).
The similarities in the formulas for f
(t) and g
(t) are a real help in
calculating λ
(t) = v (t). Since
f
(t)
2
+g
(t)
2
=
_
f (t) g
(t) −f
(t) g (t)
¸
2
g (t)
2
+
_
f (t) g
(t) −f
(t) g (t)
¸
2
f (t)
2
=
_
f (t) g
(t) −f
(t) g (t)
¸
2
146 CHAPTER VII CURVES AND ARC LENGTH
and we’ve assumed that the quantity in brackets is positive, we have
λ
(t) =
_
f
(t)
2
+g
(t)
2
= f (t) g
(t) −f
(t) g (t).
Consequently, we obtain the formulas
dx
ds
= f
(T (s)) T
(s) =
f
(t)
λ
(t)
= −g (t) = −y,
dy
ds
= g
(T (s)) T
(s) =
g
(t)
λ
(t)
= f (t) = x.
Formulas for f
(t) and g
(t) may not be nearly so simple; they depend
strongly on the speciﬁc formulas for f (t) and g (t).
It’s easy enough to imagine a curve that wraps around the circle in
ﬁnitely often, and the use of arc length to parametrize this curve should be
quite familiar. Mathematicians prefer to orient the curve in the counter
clockwise direction, with s = 0 corresponding to the point (1, 0). Then its
parametrization in terms of directed arc length is
x = cos s, y = sins for −∞ < s < ∞,
and this is taken as the deﬁnition of the circular functions cosine and
sine. For 0 < s <
π
2
, these agree with the usual trigonometric ratios
corresponding to an acute angle with radian measure s, so the terms circular
functions and trigonometric functions are used almost interchangeably. The
notation for the trigonometric functions was standardized before calculus
notation developed, which is why we usually write sin s and cos s instead
of sin (s) and cos (s). The formulas
d
ds
(cos s) = −sins and
d
ds
(sins) = cos s
are just restatements of the formulas for
dx
ds
and
dy
ds
that we derived above.
For more general curves, the derivatives
dx
ds
and
dy
ds
are used to determine
where the curve is smooth; other derivatives may not provide the right
information. For example, the graph of y = f (x) is certainly smooth
where f
(x) is continuous, but it can also be smooth where f isn’t even
differentiable; a smooth curve can have a vertical tangent line. Or when
a curve is expressed in the form x = f (t), y = g (t), the curve may not
be smooth at all points where f
(t) and g
(t) are continuous; it can have
a corner at a point where both derivatives vanish. The exercises below
illustrate some of these ideas.
EXERCISES 147
EXERCISES
7. The parametrizations of the unit circle found in the previous set of
exercises describe clockwise motion around the unit circle. By simply
computing all the indicated quantities, show that
f
(t) = g (t) λ
(t) and g
(t) = −f (t) λ
(t).
8. The graph of y = x
1/3
can be parametrized by x = t
3
, y = t for
t in the interval (−∞, ∞). Without actually expressing x and y as
functions of directed arc length, ﬁnd
dx
ds
and
dy
ds
as functions of t and
show that they are continuous.
9. The graph of x = t
3
, y = t
2
has a cusp at the origin. Show that
dy
ds
is discontinuous there by considering the cases t > 0 and t < 0
separately.
10. The graph of x = t
3
, y = t t is a smooth curve. Show that
dy
ds
is
continuous at the origin by considering the cases t > 0 and t < 0
separately.
4 THE ARCTANGENT AND ARCSINE FUNCTIONS
Although the trigonometric functions are more familiar to us, in many ways
the inverse trigonometric functions are simpler. Arc length is basically hard
to measure, and locating a point corresponding to a speciﬁed arc length is
harder still. We know exact values for the circular functions at multiples
of
π
4
, for example, because we recognize
π
4
as exactly oneeighth the
circumference of the unit circle and we can use geometric constructions
to locate the appropriate point on the unit circle. But to locate the point
(cos 1, sin1) with any degree of precision, some heavyduty calculations
are involved. Since we now do those things by just pushing a few buttons
on a calculator, it’s easy to lose sight of how much is really going on. It’s
fundamentally simpler to locate a point with given coordinates and then
determine arc length up to that point than it is to determine the coordinates
of a point corresponding to a given arc length. That makes it easier to
develop methods for calculating inverse trigonometric functions.
The simplest of the inverse trigonometric functions is the arctangent.
For P (t) the point where the line segment fromthe origin to (1, t) intersects
the unit circle, arctan (t) is the directed arc length from P (0) to P (t).
We’ll use this geometric description of P (t) to ﬁnd a formula for a curve.
Then we’ll use equation (7.4) to produce an integral for the directed arc
length.
148 CHAPTER VII CURVES AND ARC LENGTH
The formula for P (t) is simple enough:
P (t) =
_
1
√
1 +t
2
,
t
√
1 +t
2
_
.
So we call
f (t) =
1
√
1 +t
2
and g (t) =
t
√
1 +t
2
for −∞ < t < ∞,
and note that f and g are in C
1
(R).
To calculate arc length, we’ll need some derivatives. We ﬁnd
f
(t) = −
t
[1 +t
2
]
3/2
and
g
(t) =
1
[1 +t
2
]
1/2
−
t
2
[1 +t
2
]
3/2
=
1
[1 +t
2
]
3/2
.
Thus
f
(t)
2
+g
(t)
2
=
t
2
[1 +t
2
]
3
+
1
[1 +t
2
]
3
=
1
[1 +t
2
]
2
.
Consequently, our integral formula for the arctangent is
arctan (t) =
_
t
0
_
f
(τ)
2
+g
(τ)
2
dτ
=
_
t
0
dτ
1 +τ
2
for all t ∈ R.
That also proves that
d
dt
[arctan(t)] =
1
1 +t
2
,
one of the standard formulas for derivatives.
The arcsine function is slightly more complicated than the arctangent
because it is only deﬁned on [−1, 1]. Geometrically, arcsin(t) is also
deﬁned to be the directed arc length from P (0) to P (t) along the unit
circle, where P (t) is the intersection of the horizontal line y = t with the
right half of the unit circle. This time we ﬁnd that
P (t) =
_
_
1 −t
2
, t
_
.
4. THE ARCTANGENT AND ARCSINE FUNCTIONS 149
Calling
f (t) =
_
1 −t
2
and g (t) = t,
we note that
f
(t) = −
t
√
1 −t
2
for t ∈ (−1, 1).
While f
is continuous on (−1, 1), it’s not integrable over [−1, 1] so we’re
only able to write arcsin(t) as an integral for t < 1. For those t we
calculate
f
(t)
2
+g
(t)
2
=
t
2
1 −t
2
+ 1 =
1
1 −t
2
.
Hence
arcsin(t) =
_
t
0
dτ
√
1 −τ
2
for −1 < t < 1.
Since the circle is a rectiﬁable curve, we knowthat arcsin(t) is actually
continuous on [−1, 1], and we know that
arcsin(1) =
π
2
,
because the arc involved is a quartercircle. We can conclude that
lim
t→1−
__
t
0
dτ
√
1 −τ
2
_
= lim
t→1−
arcsin(t) =
π
2
.
This is often indicated in the condensed form
_
1
0
dτ
√
1 −τ
2
=
π
2
,
even though the function to be integrated is not Riemann integrable over
[0, 1]. Such expressions are called improper integrals. Even though
the notation is the same as for ordinary integrals, there are important
differences. For example, Riemann sums do not provide reliably accurate
approximations to improper integrals, no matter how small the mesh of the
partition used.
It’s also possible to calculate arctan(t) or arcsin(t) geometrically,
in much the same way as we might calculate the arc length of the entire
150 CHAPTER VII CURVES AND ARC LENGTH
circle. We can form successive approximations as lengths of brokenline
paths in a way that lets us calculate their lengths recursively. But we need
not pursue that here.
EXERCISES
11. How can our integral formula for the arctangent be used to calculate
π?
12. Show geometrically that
arcsin(t) = arctan
_
t
√
1 −t
2
_
for −1 < t < 1.
13. For t > 1, let P (t) be the ﬁrstquadrant point on the unit circle from
which the tangent line passes through (t, 0). Find the function v (t)
we integrate to ﬁnd directed arc length along this curve. What’s the
problem in using the corresponding integral to deﬁne the arcsecant?
5 THE FUNDAMENTAL TRIGONOMETRIC LIMIT
Our use of equation (7.4) and the inverse function theorem to derive the
formulas
d
ds
(sins) = cos s and
d
ds
(cos s) = −sins
is certainly not the way it’s done in introductory calculus courses. We
had the luxury of not needing the formulas quickly. They’re ordinarily
developed by using trigonometric identities to express the needed limits in
terms of the fundamental trigonometric limit
lim
s→0
sins
s
= 1.
This limit is closely related to our understanding of arc length on the
unit circle. Each side of an inscribed regular nsided polygon subtends an
angle 2π/n, so each side has length 2 sin(π/n). Consequently,
2π = lim
n→∞
2nsin(π/n),
and we can rewrite that formula as
lim
n→∞
sin(π/n)
π/n
= 1.
That’s strong evidence for the fundamental limit, but it doesn’t fully es
tablish it because every interval (−δ, δ) will contain lots of numbers that
5. THE FUNDAMENTAL TRIGONOMETRIC LIMIT 151
aren’t of the formπ/n. Just a little more work is needed. Symmetry shows
that we only need to consider s > 0, and we note that
sin[π/ (n + 1)] ≤ sin s ≤ sin(π/n) when π/ (n + 1) ≤ s ≤ π/n.
So for all such s we have
sin[π/ (n + 1)]
π/n
≤
sin s
s
≤
sin (π/n)
π/ (n + 1)
,
and we can rewrite this as
n
n + 1
·
sin [π/ (n + 1)]
π/ (n + 1)
≤
sins
s
≤
n + 1
n
·
sin(π/n)
π/n
.
Then as s decreases to 0, n → ∞and that shows (sins) /s → 1.
Figure 7.2 Comparison of s, sins, and tans.
Most calculus books establish the fundamental limit as a consequence
of the inequality
sins < s < tans for all s ∈
_
0,
π
2
_
. (7.5)
That certainly agrees with visual estimates of the distances in Figure 7.2,
where an arc of length s on the unit circle is between a vertical line of
length sins and a slanted line of length tans. To measure s we would
calculate either arcsin(sins) or arctan(tans), and the integral formulas
we developed in the previous section show that for 0 < s <
π
2
we have
sins < arcsin(sins) and arctan (tans) < tans.
Inequality (7.5) also follows from the ideas in our ﬁrst discussion of
arc length on circles, if we double all the terms by adding a bottom half
that is the mirror image of the top half. That makes 2s the length of an arc
152 CHAPTER VII CURVES AND ARC LENGTH
subtending a chord of length 2 sin s, and the ends of the arc are joined by a
brokenline path of length 2 tan s that is part of a circumscribed polygon.
Arc length is a subtle enough concept that most calculus books avoid
worrying about it too much. In the 1960s it became fashionable to establish
(7.5) by comparing areas instead. The ratio of the area of a circular sector
determined by an arc of length s to the area of the whole circle should
be the same as the ratio of s to the circumference of the circle. So since
the whole unit circle has area π, the sector must have area
1
2
s. But this
sector obviously contains a triangle of area
1
2
sins and is contained in
another triangle of area
1
2
tans, and that proves (7.5). However, this
discussion ignores the question of why the arc length and area of a circle
involve the same mysterious constant we call π. The concept of arc length
really is an important part of every geometric approach to the fundamental
trigonometric limit.
VIII
SEQUENCES AND SERIES OF
FUNCTIONS
I
n this chapter we’ll study some additional ways to produce new func
tions and learn how we can work with the functions that result. These
techniques are fairly general and have proved to be a valuable source
of functions that meet needs of engineers and scientists as well as mathe
maticians.
1 FUNCTIONS DEFINED BY LIMITS
Calculus operations usually involve limits in some sense, and the simplest
sort of limit is the limit of a sequence. In Chapter 3 we proved things about
limits of sequences of numbers. Now we’ll consider limits of sequences
of numerical functions. Generally, we’ll assume that we have an interval I
and an inﬁnite sequence {f
n
}
∞
n=1
of numerical functions whose domains
all include I. Then for each x ∈ I the sequence {f
n
(x)}
∞
n=1
is a sequence
of numbers. When each of those sequences is convergent, the formula
f (x) = lim
n→∞
f
n
(x)
153
154 CHAPTER VIII SEQUENCES AND SERIES OF FUNCTIONS
deﬁnes a numerical function f on I; we say that f is the pointwise limit
of the sequence {f
n
}
∞
n=1
. For example, the natural exponential function is
the pointwise limit of the sequence {E
n
}
∞
n=1
given by
E
n
(x) =
_
1 +
1
n
x
_
n
for −∞ < x < ∞,
and the natural logarithm is the pointwise limit of the sequence {L
n
}
∞
n=1
deﬁned by
L
n
(x) = n
_
n
√
x −1
_
for x > 0.
In Chapter 6 we used other deﬁnitions for the natural logarithm and ex
ponential function, then developed these sequences to provide useful ap
proximations. Here we’re concerned with using sequences both to deﬁne
functions and to develop their properties, and that raises a different set
of questions. How do we tell whether a given sequence of functions has
a pointwise limit? When it does, how can we establish the continuity,
differentiability, or integrability of the function it deﬁnes? How can we
calculate the derivative or the integral? Ideally, we would like a simple
way to answer such questions by inspecting the terms in the sequence used
for the deﬁnition.
Determining whether a given sequence of functions has a pointwise
limit is itself a challenge. In the case of the natural exponential function
and the natural logarithm, we used properties of those functions to explain
why the sequences {E
n
(x)}
∞
n=1
and {L
n
(x)}
∞
n=1
were convergent. But
if the pointwise limit of a sequence of functions {f
n
}
∞
n=1
is to serve as the
deﬁnition of f, then we can’t count on using properties of f or even the
value of f (x) to help show why {f
n
(x)}
∞
n=1
converges. That’s a serious
problem since the limit of a sequence is deﬁned in terms of a condition it
must satisfy instead of as the result of performing some operation on the
sequence. To get past this difﬁculty, we go back to the ideas of Chapters 1
and 3.
Recall that we ﬁrst investigated the convergence of a bounded sequence
{x
n
}
∞
n=1
of real numbers by considering the tails, deﬁning
a
m
= inf {x
n
: n ≥ m} and b
m
= sup {x
n
: n ≥ m}.
That makes {[a
m
, b
m
]}
∞
m=1
a nested sequence of closed intervals, and the
sequence {x
n
}
∞
n=1
converges if and only if b
m
− a
m
→ 0. When the
sequence does converge, the limit is the one number in
∞
m=1
[a
m
, b
m
].
But we don’t need to ﬁnd the sequence of intervals to tell whether
b
m
− a
m
→ 0; a very simple test was developed in the early nineteenth
1. FUNCTIONS DEFINED BY LIMITS 155
century by Augustin Cauchy, a French mathematician. He observed that
a sequence {x
n
}
∞
n=1
of real numbers has a limit if and only if for each
given ε > 0 there is an integer N such that x
n
−x
m
 < ε for all n, m ≥
N. Nowadays we say that sequences with this last property satisfy the
Cauchy condition. Clearly every convergent sequence satisﬁes the Cauchy
condition; when both n and m are large enough to put x
n
and x
m
within
1
2
ε of the limit, the distance between x
n
and x
m
must be less than ε. On
the other hand, we know in general that
b
N
−a
N
= diam{x
n
: n ≥ N} = sup {x
n
−x
m
 : n, m ≥ N}.
So if x
n
−x
m
 < ε for all n, m ≥ N, then 0 ≤ b
N
− a
N
≤ ε. Conse
quently, when Cauchy’s condition is satisﬁed we can prove b
m
−a
m
→ 0,
and therefore we can use {[a
m
, b
m
]}
∞
m=1
to deﬁne the limit of {x
n
}
∞
n=1
.
The fact that every sequence satisfying Cauchy’s condition is convergent
is one more version of the completeness property of R.
Of course, the pointwise limit of a sequence of functions is a somewhat
more complicated matter because we have to deal with a different sequence
{f
n
(x)}
∞
n=1
for each x. But there are no new ideas involved, just more
details to take care of. To see if the sequence has a pointwise limit,
we check the pointwise Cauchy condition; that is, we check whether it
satisﬁes the Cauchy condition at each point. Explicitly, the sequence of
numerical functions {f
n
}
∞
n=1
has a pointwise limit on I if for each x ∈ I
and each ε > 0 there is an integer N such that f
n
(x) −f
m
(x) < ε for all
n, m ≥ N. Note that in this condition, the integer N is allowed to depend
on x as well as ε; that can be important. But for now, the most important
lesson is that when {f
n
}
∞
n=1
satisﬁes the pointwise Cauchy condition on
I, then calling
f (x) = lim
n→∞
f
n
(x), x ∈ I
deﬁnes a function f on I that is the pointwise limit of the sequence.
It’s also common to use an inﬁnite series to deﬁne a function. While the
terms sequence and series are almost interchangeable in everyday speech,
they have quite different meanings in mathematics. To a mathematician,
an inﬁnite series always refers to a sum of inﬁnitely many terms. The
expression
∞
n=1
a
n
is interpreted as shorthand for the sequence {s
n
}
∞
n=1
of partial sums deﬁned by the formula s
n
=
n
k=1
a
k
. It’s sometimes
helpful to remember that the partial sums can be deﬁned recursively by
s
1
= a
1
and s
n+1
= s
n
+a
n+1
for all n ∈ N.
It’s also common to use
∞
n=1
a
n
to indicate lim
n→∞
s
n
when this se
quence converges. When we say an inﬁnite series is convergent, that’s
156 CHAPTER VIII SEQUENCES AND SERIES OF FUNCTIONS
always understood to mean that the corresponding sequence of partial
sums is convergent, so there’s little need to develop a separate theory of
functions deﬁned by inﬁnite series. Note that when
s
n
(x) =
n
k=1
a
k
(x) for all n ∈ Nand all x ∈ I,
questions about the continuity, differentiability, or integrability of the func
tions in the sequence {s
n
}
∞
n=1
can be answered by examining the functions
in
∞
n=1
a
n
one at a time.
When f is the pointwise limit of {f
n
}
∞
n=1
, it’s tempting to conjecture
that f
is the pointwise limit of {f
n
}
∞
n=1
. This is true in many important
cases; let’s look at a couple of them. For the sequence giving the natural
exponential function, we easily calculate
E
n
(x) = n
_
1 +
1
n
x
_
n−1
d
dx
_
1 +
1
n
x
_
=
_
1 +
1
n
x
_
n−1
= E
n
(x)
__
1 +
1
n
x
_
when n = −x.
Since 1 +
1
n
x → 1, we see that E
n
(x) → exp(x) for each x ∈ R, which
is the derivative of exp(x). Similarly, we can also calculate
L
n
(x) = n ·
1
n
x
1/n−1
= x
1/n−1
.
For x > 0 we have x
1/n
→ 1, so L
n
(x) → x
−1
, the derivative of log x.
For functions deﬁned by limits, making conjectures about their prop
erties is an easy matter. But verifying those conjectures is another matter
entirely. Some simple examples point up the need for supplementary hy
potheses. If we deﬁne f
n
on [0, 1] by f
n
(x) = x
n
, then it’s easy to
recognize that
lim
n→∞
f
n
(x) =
_
0, 0 ≤ x < 1
1, x = 1.
Each f
n
is continuous and strictly increasing on [0, 1], yet the pointwise
limit is constant on [0, 1) and discontinuous at 1. Or if we deﬁne g
n
on R
by
g
n
(x) =
1
n
sin n
2
x,
1. FUNCTIONS DEFINED BY LIMITS 157
then clearly g
n
(x) → 0 for each x ∈ R. In this case the pointwise limit
certainly deﬁnes a differentiable function, but since g
n
(x) = ncos n
2
x, we
see little relationship between the sequence of derivatives and the derivative
of the pointwise limit.
Now let’s look at an example that is not at all simple, a type of function
ﬁrst studied in the nineteenth century by the German mathematician Karl
Weierstrass. He found an inﬁnite series of differentiable functions converg
ing to a continuous function that isn’t differentiable at any point. That was
a real shock to the mathematics community; virtually no one had thought
such a thing was possible. It showed the need for developing the ideas of
calculus carefully, because intuition could be wrong. We’ll conclude this
section by analyzing the function deﬁned by
f (x) =
∞
n=0
4
−n
sin4
2n
x, −∞ < x < ∞.
We’ll show that there are no points at which it is differentiable and no
intervals on which it is monotonic, even though it is continuous at all
points.
First let’s see why the given series converges for all x ∈ Rand deﬁnes
a continuous function. Since the series starts with n = 0 instead of n = 1,
calling s
n
(x) the sum of the ﬁrst n terms in the series makes
s
n
(x) =
n−1
k=0
4
−k
sin4
2k
x.
Then for m < n we have
s
n
(x) −s
m
(x) =
¸
¸
¸
¸
¸
n−1
k=m
4
−k
sin4
2k
x
¸
¸
¸
¸
¸
≤
n−1
k=m
4
−k
=
1
3
_
4
1−m
−4
1−n
_
.
So we have s
n
(x) −s
m
(x) <
1
3
· 4
1−N
for all m, n ≥ N, which shows
that our series satisﬁes the pointwise Cauchy condition on all of R. More
over, the limit function f satisﬁes
f (x) −s
m
(x) ≤
1
3
· 4
1−m
for all x ∈ Rand all m ∈ N.
Now let’s prove that f is continuous at each point a ∈ R. Our task
is to show that when ε > 0 is given, there is a δ > 0 such that every
158 CHAPTER VIII SEQUENCES AND SERIES OF FUNCTIONS
x ∈ (a −δ, a +δ) satisﬁes f (x) −f (a) < ε. The key is that f is
almost the same as s
m
for any large enough value of m, and each s
m
is
continuous since it is a ﬁnite linear combination of continuous functions.
For any m, we have
f (x) −f (a) = f (x) −s
m
(x) +s
m
(x) −s
m
(a) +s
m
(a) −f (a)
≤ f (x) −s
m
(x) +s
m
(x) −s
m
(a) +s
m
(a) −f (a)
≤
1
3
· 4
1−m
+s
m
(x) −s
m
(a) +
1
3
· 4
1−m
.
So we pick mlarge enough to make 4
1−m
< ε, and since s
m
is continuous
at a there must be a δ > 0 such that every x ∈ (a −δ, a +δ) satisﬁes
s
m
(x) −s
m
(a) <
1
3
ε. That’s the δ we need.
Now we prove that f is never differentiable by using some properties
of the sine function. Since sin
_
x ±
π
2
_
= ±cos x and
(sinx + cos x)
2
+ (sinx −cos x)
2
= 2 sin
2
x + 2 cos
2
x = 2,
either sin
_
x +
π
2
_
or sin
_
x −
π
2
_
must differ from sinx by at least 1. So
with a ﬁxed and m any nonnegative integer, if 4
2m
h
m
= ±
π
2
and we
choose the sign of h
m
correctly, we get
¸
¸
sin4
2m
(a +h
m
) −sin4
2m
a
¸
¸
≥ 1.
We also have
sin4
2n
(a +h
m
) = sin4
2n
a for all n > m.
This last equation lets us write f (a +h
m
) −f (a) as the difference of two
ﬁnite sums; we regroup them as
f (a +h
m
) −f (a) = 4
−m
_
sin4
2m
(a +h
m
) −sin4
2m
a
¸
+s
m
(a +h
m
) −s
m
(a).
Hence
f (a +h
m
) −f (a) ≥ 4
−m
−s
m
(a +h
m
) −s
m
(a),
and we note that 4
−m
= 4
m
·
2
π
h
m
.
Now we bound s
m
(a +h
m
) −s
m
(a) by using the mean value the
orem. There must be a number c
m
between a and a +h
m
with
s
m
(a +h
m
) −s
m
(a) = h
m
s
m
(c
m
)
EXERCISES 159
and we calculate
¸
¸
s
m
(c
m
)
¸
¸
=
¸
¸
¸
¸
¸
m−1
k=0
4
k
cos 4
2k
c
m
¸
¸
¸
¸
¸
≤
m−1
k=0
4
k
<
1
3
· 4
m
.
So we obtain
f (a +h
m
) −f (a) > 4
m
_
2
π
−
1
3
_
h
m
.
That shows why f
(a) can’t exist. Even though h
m
→ 0 as m → ∞,
we clearly have
lim
m→∞
¸
¸
¸
¸
f (a +h
m
) −f (a)
h
m
¸
¸
¸
¸
= ∞.
Similar calculations explain why f is nowhere monotonic. Given a non
trivial interval I, when m is large enough the interval I will have a
closed subinterval I
m
of length 4
−2m
π that is centered at a point a
m
with sin4
2m
a
m
= 1. Then for each n ≥ m the term sin4
2n
x will vanish
at both ends of I
m
, and for n > mit will vanish at a
m
as well. The changes
in s
m
over I
m
are small enough to guarantee that f (a
m
) is strictly larger
than the value of f at either endpoint of I
m
.
In Chapter 2 we mentioned that we have no procedure for locating the
extrema of an arbitrary continuous function; this example gives a hint of
the difﬁculties that can be involved. When we can ﬁnd extreme values,
we usually do it by using derivatives to split the interval into subintervals
on which the function is monotonic, and then compare the values of the
function at the endpoints of the subintervals. Here that’s entirely out of the
question.
EXERCISES
1. For f
n
deﬁned on Rby the formula
f
n
(x) =
n
2
x
2
n
2
x
2
+ 1
,
let f be the pointwise limit of {f
n
}
∞
n=1
on R. Graph f. Where is f
discontinuous?
2. Deﬁne f
n
(x) = nφ(nx), where
φ(x) =
_
1 −x −1 , x −1 ≤ 1
0, x −1 > 1.
160 CHAPTER VIII SEQUENCES AND SERIES OF FUNCTIONS
Show that f
n
(x) → 0 for all x ∈ R, but that
_
2
0
f
n
(x) dx = 1 for all n.
2 CONTINUITY AND UNIFORM CONVERGENCE
Now let’s develop a general condition under which we can prove that the
pointwise limit of continuous functions is itself a continuous function.
We’ll assume that each function f
n
in the sequence {f
n
}
∞
n=1
is deﬁned
on a ﬁxed interval I, and that f
n
(x) → f (x) for all x ∈ I. How do we
prove that f is continuous at a point a ∈ I? Given ε > 0, we need to
be able to produce δ > 0 such that every x ∈ I ∩ (a −δ, a +δ) satisﬁes
f (x) −f (a) < ε. Since we usually knowmore about the functions f
n
in
the sequence than we know about the limit f, we need to use the inequality
f (x) −f (a) ≤ f (x) −f
n
(x) +f
n
(x) −f
n
(a) +f
n
(a) −f (a).
If f is the pointwise limit of the sequence, then for each x ∈ I we can ﬁnd
an n such that both
f (x) −f
n
(x) <
1
3
ε and f (a) −f
n
(a) <
1
3
ε,
and if f
n
is continuous at a then we can ﬁnd a δ > 0 such that every x in
the interval I ∩ (a −δ, a +δ) satisﬁes
f
n
(x) −f
n
(a) <
1
3
ε
as well. So what’s the problem? Doesn’t this prove that f is continuous at
a?
In fact, we’ve got a very big problem. In general we need the value
of x to tell us what n to use, and when we use the continuity of f
n
at a
to determine our δ, that means we can’t choose x until we’ve chosen n.
It’s a classic example of circular reasoning, but even if we don’t recognize
the logical fallacy we should recognize that something must be wrong
somewhere. After all, we’ve seen an example in which the pointwise limit
of a sequence of continuous functions is discontinuous.
In the previous section we used a standard method to establish the
continuity of our nowhere differentiable function. We were able to choose
m with f (x) −s
m
(x) < ε/3 for all x, rather than having to choose an
m that depended on any particular x under consideration. Choosing m
2. CONTINUITY AND UNIFORM CONVERGENCE 161
independently of x lets us use the continuity of the corresponding function
in the sequence to prove the continuity of the limit. The deﬁnition and
theorem below give a systematic way to exploit this simple idea.
Definition 2.1: Suppose that {f
n
}
∞
n=1
is a sequence of numerical
functions all deﬁned on I, and f
n
→ f pointwise on I. We say that
{f
n
}
∞
n=1
converges uniformly to f on I if for each ε > 0 there is an
integer N such that every function f
n
with n ≥ N satisﬁes
f
n
(x) −f (x) < ε for all x ∈ I.
We often indicate this condition by writing f
n
→ f uniformly or by
saying that f is the uniform limit of the sequence of functions. Note
that pointwise convergence is a consequence of the last condition in the
deﬁnition of uniform convergence, so we need not check for it separately.
Of course, when pointwise convergence fails, uniform convergence must
fail as well. Once we understand how uniform convergence differs from
pointwise convergence, it’s a simple matter to prove the theorem below.
The steps in proving it are all described above, so we’ll leave a formal
proof to the reader.
Theorem 2.1: Let {f
n
}
∞
n=1
be a sequence of numerical functions
all deﬁned on a ﬁxed interval I, and suppose that f : I →Ris the uniform
limit of {f
n
}
∞
n=1
on I. Then f is continuous at each point where all the
functions in the sequence are continuous.
One way to prove that a sequence {f
n
}
∞
n=1
of functions on an interval
I is uniformly convergent is to show that it satisﬁes a uniform Cauchy
condition on I. That is, for each ε > 0, there is an integer N such that
f
n
(x) −f
m
(x) < ε for all n, m ≥ N and for all x ∈ I. That’s the route
we took to establish uniform convergence for
∞
n=0
10
−n
sin 10
2n
x, for
example. Let’s see why that works. Of course, if a sequence satisﬁes a
uniform Cauchy condition on an interval, then it also satisﬁes a pointwise
Cauchy condition on the same interval, and so the sequence must con
verge pointwise to a limit function f. But to prove uniform convergence,
whenever ε > 0 is given we need to show there is an integer N such that
f
n
(x) −f (x) < ε for all n ≥ N and all x ∈ I. We note that for every
m, n ∈ Nand every x ∈ I we have
f
n
(x) −f (x) ≤ f
n
(x) −f
m
(x) +f
m
(x) −f (x).
162 CHAPTER VIII SEQUENCES AND SERIES OF FUNCTIONS
Thus if {f
n
}
∞
n=1
satisﬁes a uniform Cauchy condition on I, then there is
an N such that
f
n
(x) −f (x) <
ε
2
+f
m
(x) −f (x)
for all n, m ≥ N and all x ∈ I. Since f
n
→ f pointwise on I, for each
x ∈ I there must be an integer m ≥ N with f
m
(x) −f (x) < ε/2.
Therefore,
f
n
(x) −f (x) <
ε
2
+
ε
2
= ε
for all x ∈ I and all n ≥ N.
To prove that an inﬁnite series of functions is uniformly convergent,
we use the same idea but the calculations look somewhat different. Of
course, when we say
∞
n=1
f
n
(x) converges uniformly, that means that
the sequence {S
n
(x)}
∞
n=1
deﬁned by
S
n
(x) =
n
k=1
f
k
(x)
converges uniformly, and we can still show this by proving the sequence
satisﬁes a uniform Cauchy condition. But in the formula for S
n
(x) −
S
m
(x) we can always cancel the common terms in the sums, and that
gives a different look to our uniform Cauchy condition. Given ε > 0, we
need to be able to ﬁnd an integer N such that
¸
¸
¸
¸
¸
n
k=m+1
f
k
(x)
¸
¸
¸
¸
¸
< ε for all x ∈ I and all m, n with N ≤ m < n.
A moment’s thought shows that we don’t really change the usual Cauchy
condition by considering only N ≤ m < n instead of all m, n ≥ N, and
in the case of series this change simpliﬁes the notation signiﬁcantly.
In many cases we can establish the uniform convergence of an inﬁnite
series of functions by using a simple condition known as the Weierstrass
Mtest. The theorem below gives a formal statement of just what it
involves.
Theorem 2.2: Let {f
n
}
∞
n=1
be a sequence of realvalued functions
whose domains all include a given set E, and let {M
n
}
∞
n=1
be a sequence
of constants such that f
n
(x) ≤ M
n
for all n ∈ N and all x ∈ E. If
∞
n=1
M
n
is convergent, then
∞
n=1
f
n
(x) converges uniformly on E.
EXERCISES 163
Proof : There’s nothing to prove if E = ∅, and when E = ∅ we
have a simple way to choose an N such that
¸
¸
¸
¸
¸
n
k=m+1
f
k
(x)
¸
¸
¸
¸
¸
< ε for all x ∈ E and all m, n with N ≤ m < n
whenever ε > 0 is given. Calling S =
∞
n=1
M
n
, we choose N large
enough to make
¸
¸
¸S −
N
n=1
M
n
¸
¸
¸ < ε. By hypothesis,
¸
¸
¸
¸
¸
n
k=m+1
f
k
(x)
¸
¸
¸
¸
¸
≤
n
k=m+1
f
k
(x) ≤
n
k=m+1
M
k
for all x ∈ E.
Since
∞
n=1
M
n
can’t have any negative terms, we see that
n
k=m+1
M
k
≤ S −
N
k=1
M
k
< ε for N ≤ m < n.
That completes the proof.
The notion of uniform convergence is an extremely useful one, and
we’ll see other theorems in which it appears as a hypothesis. But it’s a
stronger condition than we may really need when we want to establish that a
function deﬁned as a limit is continuous at all points in its domain. To prove
continuity of a function f on an interval I, we prove continuity at each point
in I. For that we only need to consider what happens in a small subinterval
about each point. Consequently, we may be able to establish continuity
on an interval by using uniform convergence on subintervals instead of on
the whole interval. For example, if I = (−1, 1) and f
n
(x) = x
n
, then
f
n
(x) → 0 at each x ∈ I. But the convergence is not uniform on I; for
ε > 0 we see f
n
(x) ≥ ε when
n
√
ε ≤ x < 1. Thus for 0 < ε < 1 it’s not
possible to choose an n such that f
n
(x) < ε for all x ∈ I. However, for
any x
0
∈ (−1, 1), we can ﬁnd a, b with x ∈ (a, b) and [a, b] ⊂ (−1, 1).
Uniform convergence is easy to establish on the subinterval [a, b]. Note
f
n
(x) is bounded by the larger of a
n
 and b
n
 when x ∈ [a, b] and we
know both a
n
→ 0 and b
n
→ 0.
EXERCISES
3. According to the formula for the sum of an inﬁnite geometric series,
∞
n=0
x
n
=
1
1 −x
for x ∈ (−1, 1).
164 CHAPTER VIII SEQUENCES AND SERIES OF FUNCTIONS
Show that the sequence of partial sums for this series does not con
verge uniformly on (−1, 1), but that it does converge uniformly on
each interval [−a, a] with 0 < a < 1.
4. For f a given continuous function on R, let {f
n
}
∞
n=1
be the sequence
of continuous functions on Rdeﬁned by
f
n
(x) = n
_
f
_
x +
1
n
_
−f (x)
_
.
Show that {f
n
}
∞
n=1
converges pointwise if f is differentiable on R
and that the convergence is uniform if f
is also differentiable on
R with f
bounded, but is not uniform on any interval where f
is
discontinuous.
5. In Chapter 6, we showed that (1 +x/n)
n
→ e
x
for all x ∈ R. Show
that the convergence is not uniform. (An easy way is to look at
(1 +x/n)
n
−e
x
 with x = n.)
3 INTEGRALS AND DERIVATIVES
Now that we know sequences of continuous functions can be used to deﬁne
new continuous functions, we need to try to answer the usual questions we
ask about any function we encounter in calculus: Is it differentiable? Is
there a formula for its derivative? Is it integrable? Is there a formula for its
integral? In this case, we have an obvious place to look for answers: the
sequence we use to deﬁne the function.
While we often think of differentiation as simpler than integration, the
questions about integration are much easier to answer. It is surprisingly
difﬁcult to ﬁnd conditions guaranteeing that termbyterm differentiation
of a sequence of functions will produce a sequence converging to the
derivative. We’ll take a backdoor approach to that problem and tackle the
question of integration ﬁrst. Then we’ll use the fundamental theorem of
calculus to convert questions of differentiability to questions of integrating
derivatives.
Theorem 3.1: Let {f
n
}
∞
n=1
be a sequence of Riemann integrable
functions on [a, b]. If f
n
→ f uniformly, then f is also Riemann integrable
over [a, b], with
_
b
a
f (x) dx = lim
n→∞
_
b
a
f
n
(x) dx.
3. INTEGRALS AND DERIVATIVES 165
Proof : We’ll use Theorem 2.2 in Chapter 5 to prove that f is
Riemann integrable, then showthat
_
b
a
f (x) dx is the limit of the sequence
of integrals.
To apply Theorem 2.2, we must show that whenever ε > 0 is given,
there is a partition P such that the set R(P) of Riemann sums for f has
diameter less than ε. Let’s call R
n
(P) the corresponding set of Riemann
sums for f
n
. We begin by choosing an n for which
f (x) −f
n
(x) <
ε
4 (b −a)
for all x ∈ [a, b];
uniform convergence allows this. If P = {x
0
, x
1
, . . . , x
m
} is any partition
and {x
∗
1
, . . . , x
∗
m
} is any set of sampling points for P, the corresponding
elements of R(P) and R
n
(P) satisfy
¸
¸
¸
¸
¸
m
k=1
f (x
∗
k
) ∆x
k
−
m
k=1
f
n
(x
∗
k
) ∆x
k
¸
¸
¸
¸
¸
≤
m
k=1
f (x
∗
k
) −f
n
(x
∗
k
) ∆x
k
<
m
k=1
ε
4 (b −a)
∆x
k
=
1
4
ε.
Consequently, we must have
diamR(P) ≤ diamR
n
(P) +
1
2
ε,
and that is true for every partition P of [a, b]. By hypothesis, f
n
is Riemann
integrable, and so there must be a partition P with diamR
n
(P) <
1
2
ε;
that is the partition we need.
To conclude the proof, we need to show that for any ε > 0 there is an
N such that
¸
¸
¸
¸
_
b
a
f
n
(x) dx −
_
b
a
f (x) dx
¸
¸
¸
¸
< ε for all n ≥ N.
Uniform convergence makes that easy; we choose our N to make
f
n
(x) −f (x) <
ε
2 (b −a)
for all n ≥ N and all x ∈ [a, b].
With the help of Theorem 3.1, it’s easy to develop conditions that will
justify termbyterm differentiation of a convergent sequence of functions.
The theorem below gives them.
166 CHAPTER VIII SEQUENCES AND SERIES OF FUNCTIONS
Theorem 3.2: Let I be an open interval, and suppose that {f
n
}
∞
n=1
is a sequence of differentiable functions on I, with each derivative f
n
continuous on I. Suppose further that the sequence {f
n
}
∞
n=1
converges
uniformly on each closed subinterval of I, and that there is a point a ∈ I
at which {f
n
(a)}
∞
n=1
converges. Then the sequence {f
n
}
∞
n=1
converges
pointwise on I to a differentiable function f, with
f
(x) = lim
n→∞
f
n
(x) for each x ∈ I.
Proof : First we assign names to some of the limits that we know
exist. In particular, we deﬁne
f (a) = lim
n→∞
f
n
(a).
Since {f
n
(x)}
∞
n=1
must converge for each x ∈ I, we can also deﬁne a
function g on I by
g (x) = lim
n→∞
f
n
(x).
Given any x ∈ I, we can always ﬁnd an open interval (c, d) containing
x and having [c, d] ⊂ I. Since f
n
→ g uniformly on [c, d], Theorem 2.1
guarantees that g is continuous on [c, d], and in particular, g is continuous
at x.
Since g is continuous on I, we can also deﬁne a function f on I by the
formula
f (x) = f (a) +
_
x
a
g (t) dt.
That makes f differentiable on I, with
f
(x) = g (x) = lim
n→∞
f
n
(x) for all x ∈ I.
To complete the proof, we just need to show that
f (x) = lim
n→∞
f
n
(x) for all x ∈ I.
Our deﬁnitions guarantee f
n
(a) → f (a), and when x = a we may use
f
n
(x) = f
n
(a) +
_
x
a
f
n
(t) dt.
EXERCISES 167
This gives us
lim
n→∞
f
n
(x) = lim
n→∞
f
n
(a) + lim
n→∞
_
x
a
f
n
(t) dt
= f (a) +
_
x
a
g (t) dt = f (x)
by Theorem 3.1. Note that Theorem 3.1 applies directly to integration over
[a, x] when a < x, and when x < a we can apply it to integration over
[x, a] instead.
A simple example illustrates the differences between these two theo
rems very nicely. Consider
f
n
(x) =
1
n
sinn
2
x, −∞ < x < ∞.
Then f
n
→ 0 uniformly, and the limit is a function that is certainly contin
uous, integrable, and differentiable. But while
¸
¸
¸
¸
_
b
a
f
n
(x) dx −
_
b
a
0 dx
¸
¸
¸
¸
=
1
n
3
¸
¸
cos n
2
a −cos n
2
b
¸
¸
→ 0,
we see that
¸
¸
¸
¸
f
n
(x) −
d
dx
(0)
¸
¸
¸
¸
=
¸
¸
ncos n
2
x
¸
¸
,
and this can get very large as n → ∞.
EXERCISES
6. For f (x) =
∞
n=1
f
n
(x), ﬁnd conditions under which
_
b
a
f (x) dx =
∞
n=1
_
b
a
f
n
(x) dx,
and ﬁnd conditions under which
f
(x) =
∞
n=1
f
n
(x).
Note that Theorems 3.1 and 3.2 can be used for this purpose.
7. Apply the ﬁrst exercise to the geometric series to show that
−log (1 −x) =
_
x
0
dt
1 −t
=
∞
n=0
x
n+1
n + 1
for −1 < x < 1.
168 CHAPTER VIII SEQUENCES AND SERIES OF FUNCTIONS
4 TAYLOR’S THEOREM
In Chapter 4 we learned how to use linear polynomials to approximate
differentiable functions and how to use quadratic polynomials to approxi
mate functions having second derivatives. Our goal here is a general result,
known as Taylor’s theorem, that deals with using polynomials of degree n
to approximate a function on an interval I. It’s named after the English
mathematician Brook Taylor, a younger contemporary of Isaac Newton.
As we mentioned in Chapter 4, one version of Taylor’s theorem can be
derived in much the same way we derived the formula for the error in the
quadratic approximation near a point. The version we give here is less
difﬁcult to justify.
Theorem 4.1: Suppose I is an open interval, f is a function in
C
n+1
(I), and a ∈ I. Then for all x ∈ I we may write
f (x) =
n
k=0
1
k!
f
(k)
(a) (x −a)
k
+
1
n!
_
x
a
f
(n+1)
(t) (x −t)
n
dt.
The polynomial
P
n
(x) =
n
k=0
1
k!
f
(k)
(a) (x −a)
k
= f (a) +f
(a) (x −a) +· · · +
1
n!
f
(n)
(a) (x −a)
n
is called either the nth order Taylor polynomial or Taylor’s approxi
mation for f (x) near x = a. It’s the one polynomial of degree n that
satisﬁes
P
(k)
n
(a) = f
(k)
(a) for k = 0, 1, 2, . . . , n.
The last expression in Theorem4.1 is usually thought of as representing
the error when f (x) is approximated by P
n
(x). Note that if M is a number
such that
¸
¸
f
(n+1)
(t)
¸
¸
≤ M for all t in the interval between a and x, then
since (x −t)
n
doesn’t change sign on that interval the error term satisﬁes
¸
¸
¸
¸
1
n!
_
x
a
f
(n+1)
(t) (x −t)
n−1
dt
¸
¸
¸
¸
≤
¸
¸
¸
¸
M
n!
_
x
a
(x −t)
n
dt
¸
¸
¸
¸
=
M
(n + 1)!
x −a
n+1
.
4. TAYLOR’S THEOREM 169
Proof : The theorem is a consequence of the identity
d
dt
_
n
k=1
1
k!
f
(k)
(t) (x −t)
k
_
=
1
n!
f
(n+1)
(t) (x −t)
n
−f
(t), (8.1)
a formula that’s simpler than it looks. The derivative of the sum is the sum
of the derivatives, and the derivative of the kth term is
1
k!
f
(k+1)
(t) (x −t)
k
−
1
(k −1)!
f
(k)
(t) (x −t)
k−1
.
So the terms cancel in pairs when we add the derivatives, and the only two
that remain form the righthand side of (8.1). It’s also easy to prove (8.1)
by induction.
For x and a in I, we can integrate equation (8.1) since all the functions
involved are continuous. Integrating the lefthand side gives
_
x
a
d
dt
_
n
k=1
1
k!
f
(k)
(t) (x −t)
k
_
dt =
n
k=1
_
1
k!
f
(k)
(t) (x −t)
k
_
x
a
= −
n
k=1
1
k!
f
(k)
(a) (x −a)
k
,
while the righthand side yields
_
x
a
_
1
n!
f
(n+1)
(t) (x −t)
n
−f
(t)
_
dt =
1
n!
_
x
a
f
(n+1)
(t) (x −t)
n
dt
−[f (x) −f (a)].
So rearranging the terms in the integrated equation will complete the
proof.
Often we use Taylor’s approximation with n and a ﬁxed and treat x as
the only variable. But sometimes we take a different point of view. When
f is in the class C
∞
(I), we can form the approximation P
n
(x) for every n,
which deﬁnes a sequence {P
n
(x)}
∞
n=1
of approximations to f (x). Then
we may well ask whether this sequence of functions converges pointwise
to f on I, as well as whether we can use the sequence to ﬁnd integrals or
derivatives of f. Sometimes we can’t, but the cases when it does work are
important enough to form the basis for several important deﬁnitions.
Definition 4.1: When I is an open interval, a ∈ I, and f ∈
C
∞
(I), the inﬁnite series
∞
n=0
1
n!
f
(n)
(a) (x −a)
n
170 CHAPTER VIII SEQUENCES AND SERIES OF FUNCTIONS
is called the Taylor series for f (x) based at a. If there is an r > 0
such that the Taylor series for f (x) based at a converges to f (x) for all
x ∈ (a −r, a +r), we say that f is analytic at a. We call f an analytic
function if it is analytic at each point in its domain.
Many of the familiar functions of calculus turn out to be analytic
functions, but some interesting ones aren’t analytic. There are C
∞
functions
whose Taylor series based at a converge only when x = a, and there are
others whose Taylor series converge but not to f (x). We can be sure that a
function is analytic at a point if we can prove there is an interval in which
the error term approaches zero as n → ∞. But that may be quite difﬁcult
to establish unless there is a simple formula giving f
(n)
(x) for all n. Let’s
look at two examples.
The natural exponential function is an analytic function. For f (x) =
e
x
, f
(n)
(x) = e
x
for all n, so the Taylor series based at a is
∞
n=0
1
n!
e
a
(x −a)
n
.
According to Theorem 4.1, for any n we have
e
x
−
n
k=0
1
k!
e
a
(x −a)
k
=
1
n!
_
x
a
e
t
(x −t)
n
dt.
Since 0 ≤ e
t
≤ e
a+r
for all t ∈ (a −r, a +r), when x ∈ (a −r, a +r)
we have
¸
¸
¸
¸
1
n!
_
x
a
e
t
(x −t)
n
dt
¸
¸
¸
¸
≤
e
a+r
n!
¸
¸
¸
¸
_
x
a
(x −t)
n
dt
¸
¸
¸
¸
=
e
a+r
(n + 1)!
x −a
n+1
≤
e
a+r
r
n+1
(n + 1)!
.
No matter what r > 0 we consider,
lim
n→∞
e
a+r
r
n+1
(n + 1)!
= 0.
When n ≥ N ≥ 2r −1, we have
e
a+r
r
n+1
(n + 1)!
=
e
a+r
r
N
N!
·
r
N + 1
·
r
N + 2
· · ·
r
n + 1
≤
e
a+r
r
N
N!
_
1
2
_
n+1−N
→ 0 as n → ∞.
4. TAYLOR’S THEOREM 171
We’ve proved that for any x, a ∈ Rwe have
e
x
=
∞
n=0
1
n!
e
a
(x −a)
n
.
The case a = 0 is especially useful, because e
0
= 1. The formula
e
x
=
∞
n=0
1
n!
x
n
gives a useful way to calculate values for the exponential function when
x isn’t too large.
Another simple example of an analytic function is given by f (x) =
1/x; it’s analytic at every a = 0. Using induction, it’s easy to show that
f
(n)
(x) =
(−1)
n
n!
x
n+1
,
so the Taylor series for 1/x based at a is
∞
n=0
(−1)
n
a
n+1
(x −a)
n
.
According to Theorem 4.1, as long as a and x are both positive or both
negative we have
1
x
−
n
k=0
(−1)
k
a
k+1
(x −a)
k
=
1
n!
_
x
a
(−1)
n+1
(n + 1)!
t
n+2
(x −t)
n
dt
= −(n + 1)
_
x
a
_
1 −
x
t
_
n
dt
t
2
= −
1
x
_
1 −
x
t
_
n+1
¸
¸
¸
¸
x
a
=
1
x
_
1 −
x
a
_
n+1
.
When 1 −x/a < 1, the error term approaches zero as n → ∞, which
proves that
1
x
=
∞
n=0
(−1)
n
a
n+1
(x −a)
n
for all x ∈ (a −a, a +a).
We conclude this section with a caution against reading something into
the deﬁnition of analytic that isn’t there. When f is analytic and a is a
172 CHAPTER VIII SEQUENCES AND SERIES OF FUNCTIONS
point in its domain, the domain of f can be quite different from the set
of points at which the power series based at a converges to f. Here we
will not investigate the relationship between those sets; that’s a matter best
studied in another branch of mathematics called complex analysis.
EXERCISES
8. Explain why every polynomial function p (x) is analytic on R.
9. Show that sinx and cos x are analytic on R by noting that the error
terms are always
±
1
n!
_
x
a
(x −t)
n
cos t dt or ±
1
n!
_
x
a
(x −t)
n
sint dt,
and then proving that the error terms approach 0 as n → ∞.
10. The function f deﬁned by
f (x) =
_
e
−1/x
, x > 0
0, x ≤ 0
is the simplest example of a C
∞
function that is not analytic. Of
course f
(n)
(x) is zero for all x < 0, and for x > 0 it has the form
e
−1/x
P
n
(1/x) with P
n
a polynomial of degree 2n. The polynomials
can be deﬁned inductively by the rules P
0
(t) = 1 and
P
n+1
(t) = t
2
_
P
n
(t) −P
n
(t)
¸
.
Complete the argument that f ∈ C
∞
by showing that
f
(n+1)
(0) = lim
x→0
f
(n)
(x) −f
(n)
(0)
x
= 0
for all integers n ≥ 0. Why does it follow that f is not analytic at 0?
5 POWER SERIES
Any series of the form
∞
n=0
c
n
(x −a)
n
is called a power series as long as each coefﬁcient c
n
is independent of x;
we call a the base point for the series. Obviously a Taylor series based at
a will be a power series with base point a. One of our goals in this section
5. POWER SERIES 173
is to establish a sort of converse to this idea, giving a simple condition for
a power series to be the Taylor series of an analytic function. That will
help us understand how calculus operations can be performed on analytic
functions by manipulating their Taylor series.
One of the simplest power series turns out to be the most important one
to understand: the geometric series
∞
n=0
x
n
. Just about everything we
can ever learn about general power series is based on what we understand
about this one case, and we understand a lot because we have a convenient
formula for the partial sums:
n
k=0
x
k
=
1 −x
n+1
1 −x
when x = 1.
Since
¸
¸
x
n+1
¸
¸
→ 0 if x < 1 and
¸
¸
x
n+1
¸
¸
→ ∞ if x > 1, we see that the
geometric series converges for x < 1 and diverges for x > 1. The cases
x = 1 and x = −1 should be handled separately. When x = 1 we must
abandon our formula for the partial sums and use common sense instead:
n
k=0
x
k
= n when x = 1.
So at x = 1 the series diverges to ∞. At x = −1 the series also diverges,
but in a different way; the partial sums alternate between 0 and 1 instead
of either growing or approaching a limiting value. We conclude that
∞
n=0
x
n
=
1
1 −x
for x < 1
and the series diverges for x ≥ 1.
Next we ask whether termbyterm differentiation of the geometric
series produces a new series that converges to the derivative of 1/ (1 −x).
For x = 1, we can differentiate our formula for the partial sums to obtain
n
k=1
kx
k−1
=
d
dx
_
n
k=0
x
k
_
=
d
dx
_
_
1 −x
n+1
_
(1 −x)
−1
_
= −
(n + 1) x
n
1 −x
+
1 −x
n+1
(1 −x)
2
. (8.2)
174 CHAPTER VIII SEQUENCES AND SERIES OF FUNCTIONS
Consequently, when both x
n+1
→ 0 and (n + 1) x
n
→ 0 as n → ∞, we
obtain
∞
n=1
nx
n−1
= lim
n→∞
n
k=1
kx
k−1
= lim
n→∞
_
−
(n + 1) x
n
1 −x
+
1 −x
n+1
(1 −x)
2
_
=
1
(1 −x)
2
,
which is the derivative of 1/ (1 −x). We know x
n+1
→ 0 for x < 1;
we’ll show that (n + 1) x
n
→ 0 for the same values of x.
All we really need to consider is the case 0 < x < 1; the case
x = 0 presents no challenge and for x < 0 we can use (n + 1) x
n
 =
(n + 1) x
n
. For 0 < x < 1, we can drop the terms in (8.2) with minus
signs to obtain
n
k=1
kx
k−1
<
1
(1 −x)
2
for all n. Regarding x as ﬁxed, we deﬁne
M = sup
_
n
k=1
kx
k−1
: n ∈ N
_
.
Given ε > 0, there must be an N with
N
k=1
kx
k−1
> M −ε.
Since the terms in the sum are all positive, we have (n + 1) x
n
< ε for all
n ≥ N, proving that (n + 1) x
n
→ 0 for x < 1.
Now we’re ready to tackle more general power series. The theorem
below is an important step.
Theorem 5.1: At any x where the power series
∞
n=0
c
n
(x −a)
n
converges, the set of numbers {c
n
(x −a)
n
: n ∈ N} is a bounded set.
Conversely, if r is a positive number such that {c
n
r
n
: n ∈ N} is a
bounded set, then the power series
∞
n=0
c
n
(x −a)
n
converges point
wise to a differentiable function f (x) on (a −r, a +r), and the differenti
ated series
∞
n=1
nc
n
(x −a)
n−1
converges to f
(x) on the same interval.
The convergence of both series is uniform on every closed subinterval of
(a −r, a +r).
5. POWER SERIES 175
Proof : The power series converges when the sequence of partial
sums
_
n
k=0
c
k
(x −a)
k
_
∞
n=1
converges, and the terms in any convergent
sequence forma bounded set. Hence there must be a number M (depending
on x) such that all the partial sums are in [−M, M]. Therefore
c
n
(x −a)
n
 ≤ 2M
since each term is the difference of two consecutive partial sums.
For the converse, we’ll prove the last part ﬁrst and then use Theorem
3.2 to justify termbyterm differentiation of the series. Given a closed
subinterval I ⊂ (a −r, a +r), we prove that the differentiated series
converges uniformly on I by using the Weierstrass Mtest, as given in
Theorem 2.2. To use the Mtest, we need a number r
0
smaller than r but
large enough that I ⊂ [a −r
0
, a +r
0
]. Such a number exists since I is a
closed subinterval of (a −r, a +r). Then for all x ∈ I we have
¸
¸
¸kc
k
(x −a)
k−1
¸
¸
¸ ≤ k c
k
 r
k−1
0
= k
¸
¸
¸c
k
r
k−1
¸
¸
¸ (r
0
/r)
k−1
.
We’ve assumed that there is a number M with
¸
¸
c
k
r
k
¸
¸
≤ M for all k, and
so
¸
¸
¸kc
k
(x −a)
k−1
¸
¸
¸ ≤ kMr
−1
(r
0
/r)
k−1
for all x ∈ I.
Note that
∞
k=1
k (r
0
/r)
k−1
converges since 0 < r
0
/r < 1, and the
series remains convergent when we multiply by Mr
−1
. So by the Mtest,
the differentiated series converges uniformly on I. We can use the same
argument to prove that the original series converges uniformly on I, except
our bound is
¸
¸
¸c
k
(x −a)
k
¸
¸
¸ ≤ M (r
0
/r)
k
.
Of course, Theorem 3.2 applies to sequences, not series, so we must
apply it to the sequence of partial sums instead of the terms in the series.
Obviously
d
dx
_
n
k=0
c
k
(x −a)
k
_
=
n
k=1
kc
k
(x −a)
k−1
,
so the limit of the sequence of derivatives is indeed the sum of the differ
entiated series. That completes the proof.
176 CHAPTER VIII SEQUENCES AND SERIES OF FUNCTIONS
Under the hypotheses of Theorem 5.1, we can say a great deal more
about the function deﬁned by the sum of the power series. The theorem
below is a signiﬁcant step in that direction.
Theorem 5.2: If the power series
∞
n=0
c
n
(x −a)
n
converges
pointwise on (a −r, a +r) to a function f, then f is in C
∞
(a −r, a +r),
and derivatives of all orders may be found by using termbyterm differen
tiation of the power series.
Proof : All we really need to do is show that the hypotheses
guarantee that the differentiated series
∞
n=1
nc
n
(x −a)
n−1
converges to f
(x) on (a −r, a +r). Then the theorem will follow by
induction; every time we take a derivative we get a new power series that
converges on the same interval, so we can always differentiate one more
time.
Given x ∈ (a −r, a +r), we can prove that
f
(x) =
∞
n=1
nc
n
(x −a)
n−1
by using Theorem 5.1. Here, however, we don’t know that {c
n
r
n
}
∞
n=0
is
a bounded sequence, so we can’t use the theorem directly. Regarding x as
ﬁxed for the moment, we can pick r
0
with x −a < r
0
< r and apply
the theorem on (a −r
0
, a +r
0
) instead. Note that
∞
n=0
c
n
r
n
0
converges
to f (a +r
0
) by hypothesis, and the ﬁrst part of Theorem 5.1 shows that
{c
n
r
n
0
}
∞
n=1
must be a bounded sequence.
Corollary 5.1: A function deﬁned on an open interval by means
of a convergent power series is analytic at the base point, and the power
series is the Taylor series for the function at that point.
To prove the corollary, we only need to recognize that we do indeed
have power series expressions for derivatives of all orders since evaluating
any power series at the base point picks out the constant term.
The corollary can be quite useful because there are quite a few ways
to produce convergent power series. For example, replacing x by x
2
in the
power series for e
x
based at 0 shows that
f (x) = e
x
2
=
∞
n=0
1
n!
x
2n
EXERCISES 177
deﬁnes a function that is analytic at 0, with
f
(k)
(0) =
_
0, k odd
(2n)!/n!, k = 2n.
On the other hand, using elementary calculus formulas to discover a general
formula for the kth derivative of e
x
2
is difﬁcult because the number of
terms involved keeps growing.
We’ll end this section by pointing out that Theorem 5.1 can also be
used to justify termbyterm integration of a power series since each partial
sum is continuous and the sequence of partial sums converges uniformly
on each closed subinterval of (a −r, a +r).
EXERCISES
11. Prove that for any power series
∞
n=0
c
n
(x −a)
n
, the set of all x for
which it converges is an interval.
12. Prove that the power series
∞
n=0
c
n
(x −a)
n
converges for all x in
the interval (a −r, a +r) if there is a number N such that
r c
n

1/n
≤ 1 for all n ≥ N.
13. Use the example of the geometric series to show why convergence
of a power series at each point in (a −r, a +r) does not justify
termbyterm integration over [a −r, a +r].
6 TOPICS FOR FURTHER STUDY
When a power series
∞
n=0
c
n
(x −a)
n
converges for some x = a but
not for all x, there is a number R > 0 such that the series converges on
(a −R, a +R) but diverges at all points outside [a −R, a +R]. We call
Rthe radius of convergence. There are formulas for computing it fromthe
coefﬁcients, but they involve expressions that we have no effective means
to evaluate unless the coefﬁcients conform to some regular pattern.
Extensive studies of analytic functions usually deal with functions
deﬁned on a region of the complex plane rather than just along the real
line. That allows for the development of more effective methods than
we could use here. In particular, there are easy ways to show that most
elementary functions are analytic and that various combinations of analytic
functions are also analytic. This is part of the subject matter of the theory
of functions of a complex variable, the subject of many mathematics books.
Gauss, Riemann, Weierstrass, and Cauchy produced major contributions to
this branch of mathematics, and some of the world’s best mathematicians
178 CHAPTER VIII SEQUENCES AND SERIES OF FUNCTIONS
are still trying to settle a question ﬁrst raised by Riemann about the location
of the zeroes of a particular analytic function.
Power series are not the only inﬁnite series of functions that have been
studied extensively; the French mathematician Joseph Fourier introduced
series of the form
a
0
+
∞
n=1
(a
n
cos nθ +b
n
sinnθ)
to study problems of heat ﬂow. The study of such series, now called
Fourier series, has been a major part of twentieth century mathematics and
has developed into a discipline now called harmonic analysis.
Fourier series can deﬁne functions with very strange properties, such
as the example we gave of a function that is everywhere continuous but
nowhere monotonic. Prior to the time of Fourier, the sum of an inﬁnite
series of functions was thought to share any properties common to all
the terms, such as continuity and differentiability, but Fourier claimed
otherwise. The controversy could not be resolved until precise deﬁnitions
were given for all the concepts involved. At this stage it was clear that
intuition and reasoning fromsimple examples did not provide a sound basis
for calculus.
IX
ADDITIONAL COMPUTATIONAL
METHODS
I
n this ﬁnal chapter we study four independent topics that are often
introduced in ﬁrstyear calculus courses. One common thread linking
them is that each is designed to produce a numerical value by avoiding
a direct calculation of the desired quantity.
1 L’H
ˆ
OPITAL’S RULE
In calculus we often consider the limit of a quotient at a point where both the
numerator and denominator vanish. Calculating the value of a derivative
by using its deﬁnition as a limit is one such problem, for example. Of
course, we seldom calculate a derivative by working directly with the limit;
we usually have more efﬁcient ways to ﬁnd derivatives. The ease of ﬁnding
derivatives suggests investigating whether we can use them to simplify the
calculation of additional limits of this type. To start our investigation, let’s
suppose that f
1
and f
2
are numerical functions deﬁned on an interval I,
that a ∈ I with f
2
(x) = f
2
(a) for any x ∈ I except for a itself, and that
179
180 CHAPTER IX ADDITIONAL COMPUTATIONAL METHODS
we need to ﬁnd
lim
x→a
f
1
(x) −f
1
(a)
f
2
(x) −f
2
(a)
.
If we assume that f
1
and f
2
are differentiable at a, then we know there are
numerical functions g
1
and g
2
on I that are continuous at a and satisfy
and
f
1
(x) = f
1
(a) + (x −a) g
1
(x)
f
2
(x) = f
2
(a) + (x −a) g
2
(x).
Thus
f
1
(x) −f
1
(a)
f
2
(x) −f
2
(a)
=
g
1
(x)
g
2
(x)
for all x = a in I. (9.1)
If g
2
(a) = 0, then by continuity of g
1
and g
2
at a we have
lim
x→a
f
1
(x) −f
1
(a)
f
2
(x) −f
2
(a)
=
g
1
(a)
g
2
(a)
=
f
1
(a)
f
2
(a)
,
so our limit is simply the quotient of the values of the derivatives.
However, in the course of evaluating such limits, we may ﬁnd that
f
1
(a) = f
2
(a) = 0,
and then equation (9.1) isn’t much help. For example, in evaluating
lim
x→0
1 −cos x
exp(x
2
) −1
we see that the derivatives of the numerator and denominator are sinx
and 2xexp
_
x
2
_
; both vanish at x = 0. For x = 0 we have g
1
(x) =
(1 −cos x) /x and g
2
(x) =
_
exp
_
x
2
_
−1
¸
/x, and there’s no reason to
use g
1
(x) /g
2
(x) in place of the original quotient. However, once we
remember that (sinx) /x → 1, we have no trouble recognizing that
lim
x→0
f
1
(x)
f
2
(x)
= lim
x→0
sinx
2xexp(x
2
)
=
1
2
.
But is that the same as the limit of g
1
(x) /g
2
(x)? The answer is yes, as
a consequence of L’Hˆ opital’s rule, a useful tool for analyzing limits of
quotients in many cases such as this one.
This rule appeared in what is thought to be the world’s ﬁrst calculus
textbook, published in 1696 by G. F. A. de L’Hˆ opital, a French marquis.
1. L’H
ˆ
OPITAL’S RULE 181
It’s generally believed that he did not discover this formula, but that it
was discovered by John Bernoulli, a member of a remarkable family of
mathematicians and scientists. L’Hˆ opital had hired Bernoulli to teach
him mathematics, and the contract included the rights to use Bernoulli’s
discoveries. After L’Hˆ opital’s death, Bernoulli claimed the formula was
his, but L’Hˆ opital’s name remains.
The use of L’Hˆ opital’s rule is not restricted to ﬁnding limits exactly
of the form we assumed in the ﬁrst paragraph. More generally we assume
that we are looking for a limit of a quotient, and the limit can’t be found by
taking the limits of the numerator and denominator separately. In addition
to ﬁnding the limit of f (x) /g (x) at a point where both f and g vanish, it
includes limits at points where both f and g approach ∞ and limits at ∞
as well as at points in R. All the different cases can be analyzed with a few
variants of a single argument using Cauchy’s extension of the mean value
theorem, Theorem 3.3 in Chapter 4.
To set up the argument, we’ll assume that f and g are differentiable on
an open interval I, with neither g nor g
ever vanishing on I, and that the
limit we need is determined by the values of f (x) /g (x) with x ∈ I. This
framework applies directly to onesided limits and limits at ∞ or −∞;
twosided limits can be analyzed by considering lefthand and righthand
limits separately. With our assumptions, Theorem 3.3 tells us that for any
x, y ∈ I with x = y, there must be a point ξ between x and y with
f (x) −f (y)
g (x) −g (y)
=
f
(ξ)
g
(ξ)
.
That lets us write
f (x)
g (x)
=
f (y)
g (x)
+
g (x) −g (y)
g (x)
·
f (x) −f (y)
g (x) −g (y)
=
f (y)
g (x)
+
_
1 −
g (y)
g (x)
_
f
(ξ)
g
(ξ)
. (9.2)
This formula allows us to approximate values of f/g by values of f
/g
when f (y) and g (y) are small in comparison to g (x). Before we start
considering cases, let’s give a formal statement of what we’ll prove.
Theorem 1.1: Suppose that f and g are differentiable functions on
(a, ∞), with neither g nor g
ever vanishing there. Suppose further that
either
lim
x→∞
g (x) = ∞ or lim
x→∞
f (x) = lim
x→∞
g (x) = 0.
182 CHAPTER IX ADDITIONAL COMPUTATIONAL METHODS
Then if f
(x) /g
(x) has a limit at ∞ (ﬁnite or inﬁnite), f (x) /g (x) has
exactly the same limit at ∞.
Similar results hold for limits at −∞ and for lefthand or righthand
limits at a ﬁnite point when the interval (a, ∞) is replaced by an appropriate
open interval.
Proof : We’ll give the proof only for limits at ∞; our argument is
easily adapted to all the other possibilities.
First we suppose that f
/g
has the ﬁnite limit L. Then we need to
show that for every ε > 0, there is a b such that
¸
¸
¸
¸
f (x)
g (x)
−L
¸
¸
¸
¸
< ε for all x > b.
The key to this inequality is the representation for f (x) /g (x) in equation
(9.2); our hypotheses guarantee it is valid for each x, y ∈ (a, ∞) with
x = y. The idea is to make f (y) /g (x) be smaller than ε/2 and also to
make
¸
¸
¸
¸
_
1 −
g (y)
g (x)
_
f
(ξ)
g
(ξ)
−L
¸
¸
¸
¸
<
ε
2
.
We use our rules for approximate arithmetic to do this; there is a δ > 0 such
that every u and v with u −1 < δ and v −L < δ satisfy uv −L <
ε/2. We’ll choose our b so that for all x > b, there is a point y = x in
(a, ∞) such that
¸
¸
¸
¸
f (y)
g (x)
¸
¸
¸
¸
<
ε
2
and
¸
¸
¸
¸
g (y)
g (x)
¸
¸
¸
¸
< δ, (9.3)
with every ξ between x and y satisfying
¸
¸
¸
¸
f
(ξ)
g
(ξ)
−L
¸
¸
¸
¸
< δ. (9.4)
When f and g both approach 0, we can choose b ≥ a such that (9.4) is
satisﬁed for all ξ > b. Then for each x > b,
lim
y→∞
f (y)
g (x)
= lim
y→∞
g (y)
g (x)
= 0,
so any sufﬁciently large y will satisfy (9.3), and for y > x > b every value
of ξ between x and y will satisfy (9.4).
The case when g (x) → ∞requires a different treatment. We ﬁrst ﬁx
y > a such that (9.4) is satisﬁed by all ξ > y. Then since
lim
x→∞
f (y)
g (x)
= lim
x→∞
g (y)
g (x)
= 0,
EXERCISES 183
there is a b > y such that every x > b will satisfy (9.3) and will make every
value of ξ between x and y satisfy (9.4) as well.
When f
/g
has an inﬁnite limit, we have a different sort of inequality
to satisfy. Let’s assume f
(x) /g
(x) → ∞; we can treat a limit of −∞by
considering −f/g. Then our goal is to prove that for every M > 0 there is
a b ≥ a such that
f (x)
g (x)
> M for all x > b.
We still use equation (9.2), but our strategy is to make
¸
¸
¸
¸
f (y)
g (x)
¸
¸
¸
¸
< 1,
¸
¸
¸
¸
g (y)
g (x)
¸
¸
¸
¸
<
1
2
, and
f
(ξ)
g
(ξ)
> 2 (M + 1).
The rest of the argument follows the previous pattern for a ﬁnite limit.
When f and g approach 0, we use the last inequality to choose b, and then
for x > b we can ﬁnd y > x to satisfy the ﬁrst two inequalities. When g
approaches inﬁnity, we use the last inequality to choose y, and then choose
b > y so that the ﬁrst two inequalities are satisﬁed for all x > b. The reader
is encouraged to write out the details.
EXERCISES
1. Assuming that
lim
x→a
f
(x)
g
(x)
= L,
what further assumptions are needed to guarantee that
lim
x→a
f (x)
g (x)
= L?
2. Write out a proof of L’Hˆ opital’s rule for the case
lim
x→a
+
f
(x)
g
(x)
= L.
3. For f (x) = x + sinx and g (x) = x + 1, show that
lim
x→∞
f (x)
g (x)
= 1
and that f
(x) /g
(x) has no limit as x → ∞. Why doesn’t this
contradict L’Hˆ opital’s rule?
184 CHAPTER IX ADDITIONAL COMPUTATIONAL METHODS
2 NEWTON’S METHOD
From time to time we need extremely accurate numerical values for so
lutions to equations that we can’t solve algebraically. We may be able to
use the intermediate value theorem to prove that there is a solution in a
particular interval, and we can then use the bisection method to locate a
solution to within an arbitrarily high degree of accuracy. But that may take
more of an effort than we care to expend. For example, let’s consider the
equation
e
x
= 3x,
which certainly has a root in (0, 1). Consequently, performing n bisections
will reduce the problem to looking in a subinterval of length 2
−n
. To
solve the equation with 12decimalplace accuracy, we’ll need to perform
about 40 bisections, and each one requires a new evaluation of e
x
. Now
imagine that the only available way to compute e
x
is to do the calculations
by hand; it’s obvious that a more efﬁcient way of solving the equation is
needed. Fortunately, there are methods that offer the possibility of ﬁnding
rapid improvements to approximate solutions. One of the best of these is
also one of the simplest and one of the oldest and is generally known as
Newton’s method. The availability of greater computing power has made
it more important, not less, because the advantages of using Newton’s
method become more pronounced as higher levels of accuracy are sought.
It’s now a builtin program on many scientiﬁc calculators. But there are
signiﬁcant limitations on its usefulness, and to understand them we need
to examine how the method works.
We begin by writing the equation to be solved in the form
f (x) = 0,
so the solutions are xintercepts of the graph of f. For the bisection method
we require only that f be continuous on an open interval containing the
root, but for Newton’s method we need f to be differentiable. Let’s say that
r is a number such that f (r) = 0, and that we only know r approximately.
If our initial approximation to r is a, we form a new approximation by
linearizing f (x) near x = a and then ﬁnding where the graph of the
linearization crosses the xaxis. That is, we solve
f (a) +f
(a) (x −a) = 0.
Assuming f
(a) = 0, the solution is
x = a −
f (a)
f
(a)
.
2. NEWTON’S METHOD 185
We use this formula to produce a sequence of approximations. Starting
with an initial approximation x
1
, we deﬁne {x
n
}
∞
n=1
recursively by calling
x
n+1
= x
n
−
f (x
n
)
f
(x
n
)
for f
(x
n
) = 0. (9.5)
Of course, we can halt the process successfully if we ﬁnd x
n
with f (x
n
) =
0. In practice, we stop when we can no longer recognize a signiﬁcant
difference between 0 and f (x
n
). Unfortunately, if we ever encounter x
n
with f
(x
n
) = 0, the only thing we can learn from the experience is not
to try the same value of x
1
(or any of x
2
, . . . , x
n−1
either) as our initial
approximation.
Here’s how Newton’s method works for our equation
e
x
= 3x,
which we rewrite as
f (x) = e
x
−3x = 0.
Our formula becomes
x
n+1
= x
n
−
exp(x
n
) −3x
n
exp(x
n
) −3
.
Starting with x
1
= 0 and using a calculator that displays 10 digits, we ﬁnd
x
2
= 0.5
x
3
= 0.610059655
x
4
= 0.618996780
x
5
= 0.619061283
x
6
= 0.619061287,
with no further changes recorded.
To explain why Newton’s method works so well, we analyze the error
in linearizations of f. Theorem 4.1 in Chapter 4 gives us a convenient
representation for this error. When f
is differentiable on an interval (α, β)
containing both r and the approximation x
n
, we can write
0 = f (r) = f (x
n
) +f
(x
n
) (r −x
n
) +
1
2
f
(ξ
n
) (r −x
n
)
2
for some ξ
n
between r and x
n
. As long as f
(x
n
) = 0,
r −x
n
= −
f (x
n
)
f
(x
n
)
−
f
(ξ
n
)
2f
(x
n
)
(r −x
n
)
2
.
186 CHAPTER IX ADDITIONAL COMPUTATIONAL METHODS
Combining this equation with our recursion formula (9.5), we see that
r −x
n+1
= −
f
(ξ
n
)
2f
(x
n
)
(r −x
n
)
2
. (9.6)
Nowwe can pin down some assumptions that will guarantee the success
of Newton’s method. We’re searching for a solution of f (x) = 0; we
assume that there is a solution r ∈ (α, β) and that there are constants A
and B such that
¸
¸
f
(x)
¸
¸
≥ A > 0 and
¸
¸
f
(x)
¸
¸
≤ B for all x ∈ (α, β).
Then if x
n
∈ (α, β), we know that x
n+1
can be deﬁned and that
r −x
n+1
 ≤
B
2A
(r −x
n
)
2
.
There must be an ε > 0 such that
(r −ε, r +ε) ⊂ (α, β) and Bε < A,
and then for 0 < r −x
n
 < ε we have
r −x
n+1
 <
B
2A
ε r −x
n
 <
1
2
r −x
n
.
Consequently, if our initial approximation is within ε of the solution
r, then the successive errors will decrease to zero, with the error in the
next approximation roughly proportional to the square of the error in the
current one. That explains why the accuracy improves so markedly as the
approximations get closer to the solution. For the example we worked out,
the error in using x
5
was about 4 · 10
−9
. If enough decimal places could
be used, the error with x
6
would be about 2 · 10
−17
and the error with x
7
would be about 4 · 10
−34
.
Newton’s method isn’t very good for ﬁnding repeated roots of equa
tions. When f (r) = f
(r) = 0 the problem of ﬁnding r becomes quite
difﬁcult, no matter what method is used. The bisection method may fail
completely; for example, there may well be no x with f (x) < 0.
EXERCISES
4. The greatest solution of x
3
−4x
2
+ 4x −1 = 0 is r =
1
2
_
3 +
√
5
_
.
Find an ε > 0 such that when we use Newton’s method to solve
this equation, taking x
1
∈ (r −ε, r +ε) guarantees that x
2
−r <
1
2
x
1
−r.
3. SIMPSON’S RULE 187
5. What happens when we try to use Newton’s method to solve x
1/3
= 0
and start with x
1
= 0?
6. For the equation x
3
−4x
2
+4x = 0, showthat using Newton’s method
with x
n
> 2 always leads to x
n+1
−2 >
1
2
(x
n
−2).
3 SIMPSON’S RULE
Although it is theoretically possible to obtain arbitrarily accurate approx
imations to integrals by using Riemann sums for partitions with small
enough mesh size, in practice this doesn’t work. If the mesh of the partition
is decreased, the number of calculations required will increase, becoming
a source of additional error as well as aggravation. Each calculation per
formed in terms of decimals can introduce a small roundoff error, and the
cumulative effect of many small errors may not be small at all. Conse
quently, mathematicians are always looking for methods to obtain more
accuracy with fewer calculations; this is the basic problem of numerical
analysis. One of the early successes in this area is Simpson’s rule for ap
proximating integrals. The formula is named for the English mathematician
Thomas Simpson, who published it in the eighteenth century.
Calculus books often explain Simpson’s rule in terms of parabolic arcs
ﬁtted to points on the graph of the function to be integrated. However,
this explanation doesn’t lead to quantitative expressions for the error in
Simpson’s rule, but only to a general feeling that the approximation ought
to work pretty well. We’ll ﬁnd an analytic expression for the error by
deriving an identity for the integral of a numerical function f with four
continuous derivatives. The process is similar to our derivation of Taylor’s
formula in Chapter 8.
We ﬁrst note that if g is any fourthdegree polynomial with leading
coefﬁcient 1/24, then g
(4)
(x) = 1 and
d
dx
_
3
k=0
(−1)
k
g
(3−k)
(x) f
(k)
(x)
_
= f (x) −g (x) f
(4)
(x).
This lets us use the fundamental theorem of calculus to write
_
b
a
f (x) dx =
3
k=0
(−1)
k
_
g
(3−k)
(b) f
(k)
(b) −g
(3−k)
(a) f
(k)
(a)
_
+
_
b
a
g (x) f
(4)
(x) dx.
To reduce the sum, we make g and its ﬁrst derivative vanish at both
a and b; those conditions determine the remaining coefﬁcients in g. A
188 CHAPTER IX ADDITIONAL COMPUTATIONAL METHODS
convenient way to do this is to call h = b − a. Then g (x) = s
h
(x −a),
with s
h
deﬁned by
s
h
(x) =
1
24
x
2
(x −h)
2
;
note g (a) = s
h
(0) and g (b) = s
h
(h). We calculate
s
(1)
h
(x) =
1
12
x(x −h) (2x −h),
s
(2)
h
(x) =
1
12
_
(2x −h)
2
+ 2x(x −h)
_
,
s
(3)
h
(x) =
1
2
(2x −h).
So g and g
(1)
vanish at both a and b, and we also ﬁnd that
g
(2)
(a) = g
(2)
(b) =
1
12
h
2
and g
(3)
(a) = −g
(3)
(b) = −
1
2
h.
Hence our identity becomes
_
a+h
a
f (x) dx =
1
2
h[f (a) +f (a +h)]
+
1
12
h
2
_
f
(a) −f
(a +h)
¸
+
_
a+h
a
s
h
(x −a) f
(4)
(x) dx (9.7)
for all f ∈ C
4
([a, a +h]).
The last identity leads to two different formulas for
_
a+h
a−h
f (x) dx.
One comes from replacing h by 2h and a by a −h :
_
a+h
a−h
f (x) dx = h[f (a −h) +f (a +h)]
+
1
3
h
2
_
f
(a −h) −f
(a +h)
¸
+
_
a+h
a−h
s
2h
(x −a +h) f
(4)
(x) dx. (9.8)
Or we can use (9.7) to integrate over [a −h, a] and [a, a +h] separately
and add the resulting equations. To simplify the addition of the resulting
3. SIMPSON’S RULE 189
integrals, we unify the formulas for the integrands. On [a −h, a], we have
s
h
(x −(a −h)) =
1
24
(x −a +h)
2
(x −a)
2
= s
h
(x −a),
and that agrees with s
h
(x −a) on [a, a +h]. So we ﬁnd
_
a+h
a−h
f (x) dx =
1
2
h[f (a −h) + 2f (a) +f (a +h)]
+
1
12
h
2
_
f
(a −h) −f
(a +h)
¸
+
_
a+h
a−h
s
h
(x −a) f
(4)
(x) dx. (9.9)
We get a still more useful formula for
_
a+h
a−h
f (x) dx by subtracting
1/3 of (9.8) from 4/3 of (9.9) to make the h
2
terms cancel:
_
a+h
a−h
f (x) dx =
1
3
h[f (a −h) + 4f (a) +f (a +h)]
+
_
a+h
a−h
K
h
(x −a) f
(4)
(x) dx, (9.10)
where we’ve called
K
h
(x) =
4
3
s
h
(x) −
1
3
s
2h
(x +h).
Equation (9.10) is the formula we need to analyze Simpson’s rule.
The ﬁrst term on the right of equation (9.10) is Simpson’s approxima
tion to
_
a+h
a−h
f (x) dx. The second term is the error in the approximation
when f is in C
4
[a −h, a +h]. Obviously there is no error at all when f is
a polynomial of degree three or less. To understand the error in general, we
need to examine K
h
(x −a) on the interval [a −h, a +h], and we might
as well take a = 0. Since
s
2h
(x +h) =
1
24
(x +h)
2
(x −h)
2
=
1
24
_
x
2
−h
2
_
2
,
we see that K
h
(x) is an even function, and for x ≥ 0 we have
K
h
(x) =
4
72
x
2
(x −h)
2
−
1
72
(x +h)
2
(x −h)
2
=
1
72
(x −h)
2
_
4x
2
−(x +h)
2
_
.
190 CHAPTER IX ADDITIONAL COMPUTATIONAL METHODS
Since
4x
2
−(x +h)
2
= (3x +h) (x −h) = 3 (x −h)
2
+ 4h(x −h),
we see that K
h
(x) ≤ 0 on [0, h] with
_
h
0
K
h
(x) dx =
1
72
_
h
0
_
3 (x −h)
4
+ 4h(x −h)
3
_
dx
=
1
72
_
3
5
h
5
−h
5
_
= −
1
180
h
5
.
By symmetry, K
h
(x) ≤ 0 on [−h, h] as well, and
_
h
−h
K
h
(x) dx = −
1
90
h
5
.
Since K
h
(x −a) doesn’t change sign on [a −h, a +h], when f
(4)
is
continuous there the mean value theorem for integrals lets us express the
error conveniently as
_
a+h
a−h
K
h
(x −a) f
(4)
(x) dx = −
1
90
h
5
f
(4)
(ξ)
for some ξ in the interval [a −h, a +h].
Since the error in using Simpson’s approximation for integration over
the interval [a −h, a +h] seems to be on the order of h
5
, we don’t want to
use it directly for integration over long intervals. A much better procedure
is to subdivide an interval [a, b] into shorter subintervals, use Simpson’s
approximation on each, and then add the results. The easiest thing to do is
to use equal subintervals; just pick an integer n and call
h =
1
2n
(b −a) and x
k
= a +kh for k = 0, 1, . . . , 2n.
Since [x
2k−2
, x
2k
] = [x
2k−1
−h, x
2k−1
+h], we have
_
b
a
f (x) dx =
n
k=1
_
x
2k
x
2k−2
f (x) dx
=
n
k=1
1
3
h[f (x
2k−2
) + 4f (x
2k−1
) +f (x
2k
)]
−
1
90
n
k=1
h
5
f
(4)
(ξ
k
).
EXERCISES 191
The ﬁrst of these last two sums can be regrouped as the form of
Simpson’s approximation most commonly used:
1
3
h[f (x
0
) + 4f (x
1
) + 2f (x
2
) + 4f (x
3
) + 2f (x
4
) +· · ·
· · · + 2f (x
2n−2
) + 4f (x
2n−1
) +f (x
2n
)].
The second one represents the error; let’s ﬁnd a simpler form for it. We’ve
assumed f
(4)
to be continuous on [a, b], so the intermediate value theorem
guarantees that the average of any ﬁnite set of its values on [a, b] is another
value. Hence the error is
−
n−1
k=0
1
90
h
5
f
(4)
(ξ
k
) = −
n
90
h
5
f
(4)
(ξ) = −
b −a
180
h
4
f
(4)
(ξ)
for some ξ ∈ [a, b]. That’s the usual representation of the error in Simpson’s
rule.
EXERCISES
7. Check the last formula for the error in Simpson’s rule by comparing
the exact value of
_
2h
0
x
4
dx to Simpson’s approximation.
8. Simpson’s rule can still give a useful approximation when the function
to be integrated isn’t in C
4
, but our representation for the error is no
longer valid. Show that when Simpson’s rule is used to approximate
_
1
0
√
xdx, the error is always at least
_√
2 −
4
3
_
h
3/2
. Examining
_
2h
0
√
xdx and
_
1
2h
√
xdx separately can help since
√
x ∈ C
4
[2h, 1].
9. Approximate log x for x > 1 by using Simpson’s rule with h =
1
2
(x −1) to approximate the integral deﬁning it, and ﬁnd upper and
lower bounds for the error in this approximation when 1 < x ≤ 3.
10. Use the approximation from the previous problem and Newton’s
method to approximate e, the solution of log x = 1. Find an al
gebraic equation whose solution should give a better estimate.
4 THE SUBSTITUTION RULE FOR INTEGRALS
Typically, students in calculus ﬁrst encounter substitutions as an aid to
ﬁnding antiderivatives:
_
f (g (x)) g
(x) dx =
_
f (u) du with u = g (x).
This is simply a help in keeping track of the effect of the chain rule. When
we identify a function F with F
(u) = f (u), we can then appeal to the
192 CHAPTER IX ADDITIONAL COMPUTATIONAL METHODS
fundamental theorem of calculus:
_
b
a
f (g (x)) g
(x) dx = F (g (x))
b
a
= F (g (b)) −F (g (a))
The last quantity is recognized as
_
g(b)
g(a)
f (u) du, and that’s the origin of the
substitution rule for integrals. But there are some potential difﬁculties that
this approach ignores, and they can come back to haunt us. For example,
when we start wondering whether
_
b
a
f (g (x)) g
(x) dxand
_
g(b)
g(a)
f (u) du
are equally amenable to numerical approximation, we discover that there
can be some real surprises.
Here’s an example involving a simple but unfortunate substitution. For
f (u) = u
−1/4
and g (x) = x
4/3
, we see easily that
_
8
0
f (g (x)) g
(x) dx =
_
8
0
_
x
4/3
_
−1/4
·
4
3
x
1/3
dx
=
4
3
_
8
0
dx =
32
3
.
But
_
g(8)
g(0)
f (u) du =
_
16
0
u
−1/4
du,
and u
−1/4
is not Riemann integrable over [0, 16]. Trying to apply Simpson’s
rule to
_
16
0
u
−1/4
du quickly reminds us why such integrals are called
improper and why that distinction is important.
Now that we’re aware that integrability may be a problem, we’re ready
to state and prove a simple version of the substitution theorem. We’ll prove
it by using Theorem 2.1 in Chapter 5.
Theorem 4.1: Let g be a continuous, increasing function on [a, b],
with g differentiable everywhere in (a, b) and g
Riemann integrable over
[a, b]. Suppose also that f is Riemann integrable over [g (a), g (b)]. Then
f (g (x)) g
(x) deﬁnes a Riemann integrable function over [a, b], and
_
b
a
f (g (x)) g
(x) dx =
_
g(b)
g(a)
f (u) du.
Proof : We’ll need to partition [a, b] and [g (a), g (b)] simultane
ously, and the properties of g make this easy. For P = {x
0
, x
1
, . . . , x
n
}
a partition of [a, b], we’ll call P
= {u
0
, u
1
, . . . , u
n
} the partition of
4. THE SUBSTITUTION RULE FOR INTEGRALS 193
[g (a), g (b)] with u
k
= g (x
k
) for k = 0, 1, 2, . . . , n. Since g is con
tinuous and monotonic on [a, b], every partition of [g (a), g (b)] can be
expressed as P
for some partition P of [a, b]. Given arbitrary sampling
points {x
∗
k
}
n
k=1
for P, we call u
∗
k
= g (x
∗
k
). Since g is monotonic the set
{u
∗
k
}
n
k=1
will be a set of sampling points for P
.
The key to the proof is the mean value theorem; there is at least one
point ξ
k
in each interval (x
k−1
, x
k
) such that
∆u
k
= g (x
k
) −g (x
k−1
) = g
(ξ
k
) ∆x
k
.
Consequently,
f (g (x
∗
k
)) g
(x
∗
k
) ∆x
k
−f (u
∗
k
) ∆u
k
= f (u
∗
k
)
_
g
(x
∗
k
) −g
(ξ
k
)
¸
∆x
k
for each k. Transposing the second term and summing from k = 1 to n,
we obtain
n
k=1
f (g (x
∗
k
)) g
(x
∗
k
) ∆x
k
=
n
k=1
f (u
∗
k
)
_
g
(x
∗
k
) −g
(ξ
k
)
¸
∆x
k
+
n
k=1
f (u
∗
k
) ∆u
k
. (9.11)
Equation (9.11) points the way to proving the substitution theorem
since the sum on the left represents an arbitrary Riemann sum for (f ◦ g)g
associated with P and the last sum on the right is a Riemann sum for f
associated with P
. Our strategy is to show that for any ε > 0, by choosing
P carefully we can guarantee that the ﬁrst sum on the right of (9.11) is
smaller than
1
2
ε and that the last sum is within
1
2
ε of
_
g(b)
g(a)
f (u) du.
Since f is Riemann integrable over [g (a), g (b)] by hypothesis, when
ε > 0 is given we know there is a partition P
0
of [g (a), g (b)] with
R
_
P
0
_
⊂
_
_
g(b)
g(a)
f (u) du −
1
2
ε,
_
g(b)
g(a)
f (u) du +
1
2
ε
_
.
Furthermore, every Riemann sum associated with any reﬁnement of P
0
must be in the same interval. The partition P
0
corresponds to a partition
P
0
of [a, b], and choosing P to be a reﬁnement of P
0
will make P
be a
reﬁnement of P
0
.
To ﬁnd out how to choose P, we call m
k
and M
k
the inﬁmum and
supremum of the values of g
over the subinterval [x
k−1
, x
k
], and choose a
number M with f (u) ≤ M for all u ∈ [g (a), g (b)]. Then we have
¸
¸
f (u
∗
k
)
_
g
(x
∗
k
) −g
(ξ
k
)
¸
∆x
k
¸
¸
≤ M [M
k
−m
k
] ∆x
k
,
194 CHAPTER IX ADDITIONAL COMPUTATIONAL METHODS
so
¸
¸
¸
¸
¸
n
k=1
f (u
∗
k
)
_
g
(x
∗
k
) −g
(ξ
k
)
¸
∆x
k
¸
¸
¸
¸
¸
≤
n
k=1
M [M
k
−m
k
] ∆x
k
= M [U (P) −L(P)],
where U (P) and L(P) are the upper and lower sums for g
(x). Since
g
is Riemann integrable by hypothesis, we can choose a reﬁnement P of
P
0
to make this last expression be smaller than
1
2
ε, and that completes the
proof.
It’s not hard to extend the substitution theorem somewhat. With just
a slight change in our arguments, we can also prove a version where g is
decreasing instead of increasing. And since we know how to add integrals
over adjacent subintervals, we can also extend it to any case where we can
split [a, b] into ﬁnitely many subintervals, with all the needed hypotheses
satisﬁed on each. Accordingly, we state a more general version of the
substitution theorem below; we won’t provide any further arguments for
its validity.
Theorem 4.2: Let g be a continuous function on [a, b], and suppose
that g
(x) exists at all but ﬁnitely many points in [a, b], agreeing with a
Riemann integrable function at the points where it is deﬁned. Suppose also
that there is a partition P = {x
0
, . . . , x
n
} of [a, b] such that g is monotonic
on each subinterval [x
k−1
, x
k
]. If f is Riemann integrable over the image
of [a, b] under g, then f (g (x)) g
(x) is Riemann integrable over [a, b] and
_
b
a
f (g (x)) g
(x) dx =
_
g(b)
g(a)
f (u) du.
The important thing to notice about this more general theorem is that
simply assuming the integrability of f over [g (a), g (b)] is not enough; we
need integrability over the entire image of [a, b]. For example, consider
_
b
a
f (g (x)) g
(x) dx =
_
2
−1
_
1
√
x
2
+ 1 −1
_
xdx
√
x
2
+ 1
.
Here g (x) =
√
x
2
+ 1, so that [g (a), g (b)] =
_√
2,
√
5
¸
, and f (u) =
(u −1)
−1
is Riemann integrable over that interval. However, the image of
[−1, 2] under g is
_
1,
√
5
¸
, and f is not integrable over that interval. And in
fact, f (g (x)) g
(x) is not integrable over [−1, 2]; it’s not even considered
improperly integrable.
EXERCISES 195
In summary, we must be careful when making substitutions in deﬁnite
integrals. Unlike matter and energy, integrability can easily be created or
destroyed.
EXERCISES
11. The substitution theorem can be proved with slightly different as
sumptions. Instead of assuming that g is continuous on [a, b] and
differentiable on (a, b) with its derivative g
Riemann integrable, we
can assume that g
is a given Riemann integrable function and that g
satisﬁes
g (x) = g (a) +
_
b
a
g
(t) dt for a ≤ x ≤ b.
Assumptions about the monotonicity of g are then conveniently given
in terms of the sign of g
. Assume that g
> 0 on [a, b] and prove the
substitution theorem with these modiﬁed hypotheses.
12. Show that if f is Riemann integrable over [−1, 1], then
_
b
a
f (sinx) cos xdx =
_
sin b
sin a
f (u) du for every a, b ∈ R.
13. What assumptions are needed to obtain
_
b
a
f (x) dx =
_
arcsin(b)
arcsin(a)
f (sint) cos t dt?
14. Suppose that g
(x) is positive and continuous on [a, b] and that f is
not Riemann integrable over the entire interval [g (a), g (b)] but only
over all subintervals [g (t), g (b)] with t ∈ (a, b). Prove that
lim
t→a+
_
b
t
f (g (x)) g
(x) dx = lim
s→g(a)+
_
g(b)
s
f (u) du
in the sense that if either limit exists, then so does the other and the
limits are equal.
This Page Intentionally Left Blank
REFERENCES
1. R. P. Boas, A Primer of Real Functions, 3rd ed., Providence, RI:
Mathematical Association of America, 1994.
2. E. Landau, Foundations of Analysis, New York: Chelsea Publishing
Co., 1951.
3. G. F. Simmons, Calculus withAnalytic Geometry, NewYork: McGraw
Hill Book Company, 1985.
4. C. B. Boyer, A History of Mathematics, NewYork: John Wiley &Sons,
1968.
5. W. Rudin, Principles of Mathematical Analysis, 3rd ed., New York:
McGrawHill Book Company, 1976.
197
I NDEX
absolute value, 36, 102
analytic function, 170, 176
antiderivative, 97, 108
approximate arithmetic, 18, 30,
48, 51
Archimedes, 135
arc length, 134, 136, 139
arclength function, 137
arcsecant, 150
arcsine, 148, 149
arctangent, 147, 148
area, 92
axiom, 7, 13
base, 120, 121
base point, 172
Bernoulli, John, 181
binomial coefﬁcient, 78
binomial theorem, 78
bisection, 9
bounded function, 55, 98
bounded interval, 7
bounded sequence, 47
broken line, 133
C
n
(I), 76
Cauchy, Augustin, 70, 155
Cauchy condition, 155
Cauchy’s meanvalue theorem, 70,
73
chain rule, 65
circular function, 146
circumference, 135
closed interval, 7, 22
closed set, 22
combining functions, 27, 30, 51,
53, 62
combining sequences, 48, 49
common extension, 37
completeness property, 5, 8, 11,
13, 155
complex analysis, 177
composition, 27, 30
compound interest, 131
198
INDEX 199
continuity, 28, 33
continuous compounding, 131
continuous function, 29, 59, 85,
163
convergent sequence, 47
critical point, 59, 67
curve, 132
decimal expansion, 4, 5
decision rule, 31, 36
Dedekind cuts, 22
deﬁnite integral, 97
degenerate interval, 6
derivative, 59
diameter, 14
differentiable function, 58, 59, 69,
82, 86, 166
differential equation, 111
differentiation, 60
directed arc length, 143
discontinuity, 32, 36, 38
discontinuous, 32
divergent sequence, 48
domain, 24, 26
empty set, 2
endpoint, 6, 10, 11
error term, 131, 168, 189, 191
exact arithmetic, 17
exponential function, 119, 129
exponential notation, 117
exp(x), see natural exponential
function
extreme value, 41, 59, 68, 159
extreme value theorem, 41
f : A → B, see function
f ◦ g, see composition
F
1
(a, b), F
2
(a, b), see partial deriva
tive
ﬁnite interval, 7
Fourier, Joseph, 178
Fourier series, 178
function, 24
fundamental theoremof calculus,
112, 164
generalized derivative, 111
geometric series, 164, 167, 173
graph, 26
greatest element, 11
greatest lower bound, 12
harmonic analysis, 178
hyperbolic cosine, 129
hyperbolic sine, 129
image, 24
implicit differentiation, 84
implicit function theorem, 89
improper integral, 149, 192
indeﬁnite integral, 97
index set, 3
indexed collection, 3
induction, 2, 3, 5
inductive deﬁnition, 2, 5
inf, see inﬁmum
inﬁmum, 12, 15
inﬁnite interval, 7
inﬁnite limit, 55
inﬁnite series, 155
inﬁnitely differentiable, 76, 169
integers, 3
integration, 92
integration by parts, 115
intermediate value, 35
intermediate value theorem, 33,
44, 69
intersection, 2, 8, 14, 101
interval, 6, 33, 36, 69
inverse function, 79
inverse function theorem, 81, 127,
144
200 INDEX
inverse trigonometric function, 147
jump discontinuity, 53
L(P), see lower sum
L’Hˆ opital, G. F. A., 180
L’Hˆ opital’s rule, 70, 180
λ(t), see arclength function
law of exponents, 117, 128
law of logarithms, 118
least element, 3, 11
least upper bound, 12
lefthand limit, 52
Leibniz, Gottfried, 61
Leibniz’s formula, 115
Leibniz’s notation, 61, 76
Leibniz’s rule, 77
limit, 46
limit at ∞, 53
limit of a function, 49
limit of a sequence, 47
linear combination, 62, 95
linearization, 58, 91
logarithm, 118
logarithmic function, 119, 128
log (x), see natural logarithm
lower bound, 13
lower sum, 100
Mtest, see Weierstrass Mtest
maximum, 41
mean value, 66
mean value theorem, 67, 69, 70,
96
mean value theoremfor integrals,
106
mesh, 93, 99
midpoint, 13, 34
minimum, 41
monotonic, 103, 139
N, see natural numbers
natural exponential function, 127,
130, 154, 170
natural logarithm, 124, 154
natural numbers, 1
nested sequence, 8
nested sequence of sets, 34, 46
Newton, Isaac, 61
Newton’s method, 184
nondegenerate interval, 9
nonterminating decimal, 8, 9
nowhere differentiable function,
157
nowhere monotonic function, 157
∅, see empty set
numerical analysis, 187
numerical function, 26
open interval, 7, 10, 22
open set, 22
operation, 4, 24
P, see partition
parameter, 25, 133
parametric representation, 133, 144
partial derivative, 85, 115
partial sum, 155
partition, 93, 133
Pascal’s triangle, 78
Peano, Giuseppe, 2
point of continuity, 29, 38
pointwise Cauchy condition, 155
pointwise convergence, 161
pointwise limit, 153, 160
polar coordinates, 27
polynomial, 54, 58
power series, 172
predecessor, 3, 4
product rule, 63, 77
Q, see rational numbers
quadratic approximation, 72, 91,
131
INDEX 201
quotient rule, 64
R, see real numbers
R(P), see set of Riemann sums
radius of convergence, 177
range, 24, 26
rational numbers, 4
real numbers, 5
rectiﬁable curve, 134, 135, 140
recursive deﬁnition, 5
reduced partition, 99
reﬁnement, 99, 101
relative error, 22, 131
removable discontinuity, 39, 49,
50, 58
repeated root, 76, 186
repeating decimal, 4
restriction, 39
Riemann, Bernhard, 93
Riemann integrable, 94, 98, 101,
104, 140, 164, 192
Riemann integral, 100
Riemann sum, 94, 99, 108, 124
righthand limit, 52
Rolle’s theorem, 67, 71
Σnotation, 5
sampling point, 93
scientiﬁc method, 25
second derivative, 73
second derivative test, 74
segment of a curve, 133
sequence, 3, 46
set of Riemann sums, 98
Simpson, Thomas, 187
Simpson’s rule, 189, 191
slope, 60
smooth, 146
spline function, 38
step function, 107
substitution formula, 191, 194
successive points on a curve, 134
successor, 1, 3, 4
sup, see supremum
supremum, 12, 15
tail of a sequence, 46, 47
tangent line, 60
Taylor, Brook, 168
Taylor polynomial, 168
Taylor series, 170
Taylor’s theorem, 74, 168
termbytermdifferentiation, 164,
174
termbytermintegration, 164, 177
terminating decimal, 4
transcendental function, 116
trigonometric function, 146, 150
trivial exponential function, 120
trivial logarithmic function, 120
U (P), see upper sum
uniform Cauchy condition, 161
uniform continuity, 104
uniform convergence, 161, 162
uniform limit, 161
uniformly continuous, 43
upper bound, 12
upper sum, 100
v (t), see velocity
velocity, 67, 139
Weierstrass, Karl, 157
Weierstrass Mtest, 162, 175
wellordered set, 2
Z, see integers
This Page Intentionally Left Blank
This action might not be possible to undo. Are you sure you want to continue?
Use one of your book credits to continue reading from where you left off, or restart the preview.